diff --git a/.github/workflows/agent-pr-review.yaml b/.github/workflows/agent-pr-review.yaml
index d08f6e094f..34b7d05168 100644
--- a/.github/workflows/agent-pr-review.yaml
+++ b/.github/workflows/agent-pr-review.yaml
@@ -43,22 +43,22 @@ jobs:
           ai_base_url: http://litellm.ai.svc.cluster.local/v1
           ai_api_key: ${{ secrets.LITELLM_API_KEY }}
           # omniroute routes to a larger free-tier model than the self-hosted 27B
-          # (see litellm routerSettings.fallbacks); falls back to qwen-3.6-fast
+          # (see litellm routerSettings.fallbacks); falls back to qwen-3.8-fast
           # automatically if the free provider fails or times out.
           ai_model: omniroute
           # v2.1.10 inherits ai_fallback_base_url from ai_base_url when unset, then
           # requires ai_fallback_model whenever ai_fallback_base_url is set (even
           # inherited) — fails fast with no model call otherwise. Point this at
-          # qwen-3.6-fast directly (same litellm endpoint) rather than duplicating
+          # qwen-3.8-fast directly (same litellm endpoint) rather than duplicating
           # ai_model: litellm's routerSettings.fallbacks already retries omniroute
           # failures internally, so this only fires if that whole chain fails —
           # a real second attempt via a distinct model, not a no-op retry.
-          ai_fallback_model: qwen-3.6-fast
-          # omniroute's proxied models don't share SGLang's sampling_defaults, so
+          ai_fallback_model: qwen-3.8-fast
+          # omniroute's proxied models don't share the self-hosted model's defaults, so
           # an empty temperature has no known-good fallback here — pin explicitly.
           ai_temperature: "0.2"
           ai_response_format: json_schema
-          # Covers both omniroute (observed 2-74s) and, on fallback, qwen-3.6-fast's
+          # Covers both omniroute (observed 2-74s) and, on fallback, qwen-3.8-fast's
           # ~10 tok/s cold prefill on this GPU (155-220s near ~100K ctx, validated).
           ai_request_timeout_sec: "600"
           verdict_policy: findings_severity_gated
diff --git a/.renovaterc.json5 b/.renovaterc.json5
index 70f60ca0ef..e7212be493 100644
--- a/.renovaterc.json5
+++ b/.renovaterc.json5
@@ -42,8 +42,12 @@
       description: "HuggingFace model files revision-pinned in llmkube Model sources (HF model repos are git repos; digest = repo commit sha)",
       customType: "regex",
       managerFilePatterns: ["/kubernetes/apps/ai/llmkube/models/.+\\.ya?ml$/"],
+      // Two source forms in use: the resolve/ URL for single-file GGUFs, and
+      // hf://repo@sha for multi-file repos (qwen38, muse-glimmer). Without the
+      // second, the primary serving model's weights digest tracks nothing.
       matchStrings: [
-        "source:\\s+https://huggingface\\.co/(?<depName>[^/\\s]+/[^/\\s]+)/resolve/(?<currentDigest>[a-f0-9]{40})/"
+        "source:\\s+https://huggingface\\.co/(?<depName>[^/\\s]+/[^/\\s]+)/resolve/(?<currentDigest>[a-f0-9]{40})/",
+        "source:\\s+hf://(?<depName>[^/\\s]+/[^/\\s]+)@(?<currentDigest>[a-f0-9]{40})"
       ],
       packageNameTemplate: "https://huggingface.co/{{depName}}",
       currentValueTemplate: "main",
@@ -52,7 +56,9 @@
     {
       description: "opencode npm plugins (executed in-process in an agent pod, so they stay pinned)",
       customType: "regex",
-      managerFilePatterns: ["/kubernetes/apps/ai/opencode/app/config/opencode\\.jsonc$/"],
+      managerFilePatterns: [
+        "/kubernetes/apps/ai/opencode/app/config/opencode\\.jsonc$/"
+      ],
       matchStrings: [
         '"(?<depName>@?[\\w.-]+(?:/[\\w.-]+)?)@(?<currentValue>\\d+\\.\\d+\\.\\d+)"'
       ],
@@ -70,12 +76,6 @@
       automergeType: "pr"
     },
     // Grouping rules
-    {
-      description: "sglang-rdna4 deployed image digest — built by our fork of the upstream RDNA4 image, not a pipeline in this repo; review before the HelmRelease rolls the serving pod (Recreate on the single GPU)",
-      matchDatasources: ["docker"],
-      matchPackageNames: ["ghcr.io/tanguille/sglang-rdna4"],
-      automerge: false
-    },
     {
       description: "Actions Runner Controller Group",
       groupName: "actions-runner-controller",
@@ -99,7 +99,7 @@
       matchDatasources: ["docker", "github-releases"],
       // matches all siderolabs/* packages (talos, installer, etc.); both slashes are regex
       // delimiters, so the inner one must stay escaped AND closed or config validation fails
-      matchPackageNames: ["/siderolabs\\//"],
+      matchPackageNames: ["/siderolabs\\//"]
     },
     {
       description: "Rook-Ceph Group",
diff --git a/docs/llm-hosting/bench/concsweep.py b/docs/llm-hosting/bench/concsweep.py
index 61b81dfa17..5a03d9ff7e 100644
--- a/docs/llm-hosting/bench/concsweep.py
+++ b/docs/llm-hosting/bench/concsweep.py
@@ -17,7 +17,9 @@ def _port():
 
 
 URL = f"http://127.0.0.1:{_port()}/v1/completions"
-MODEL = "qwen-3.6"
+# 2nd arg overrides the served-model-name, since the port-forward target decides
+# which engine answers. Hardcoding it once cost a whole run of silent 0.00 tok/s.
+MODEL = sys.argv[2] if len(sys.argv) > 2 else "qwen-3.8"
 GEN = 96
 
 
diff --git a/docs/llm-hosting/bench/spectest.py b/docs/llm-hosting/bench/spectest.py
index 82949545af..d63ff1d396 100644
--- a/docs/llm-hosting/bench/spectest.py
+++ b/docs/llm-hosting/bench/spectest.py
@@ -14,6 +14,9 @@ def _port():
 
 
 URL = f"http://127.0.0.1:{_port()}/v1/completions"
+# 2nd arg overrides the served-model-name, since the port-forward target decides
+# which engine answers. Hardcoding it once cost a whole run of silent 0.00 tok/s.
+MODEL = sys.argv[2] if len(sys.argv) > 2 else "qwen-3.8"
 CODE = "\n".join(f"def handler_{i}(request, context):\n"
                  f"    payload = request.get('payload_{i}')\n"
                  f"    if payload is None:\n"
@@ -21,7 +24,7 @@ def _port():
                  f"    return {{'status': 200, 'body': payload}}\n" for i in range(240))
 prompt = (f"Here is a Python module:\n\n{CODE}\n\n"
           "Reproduce handler_0 through handler_12 EXACTLY as written above, verbatim:\n\n")
-b = json.dumps({"model": "qwen-3.6", "prompt": prompt, "max_tokens": 400,
+b = json.dumps({"model": MODEL, "prompt": prompt, "max_tokens": 400,
                 "temperature": 0, "ignore_eos": True}).encode()
 t0 = time.perf_counter()
 rq = urllib.request.Request(URL, data=b, headers={"Content-Type": "application/json"})
diff --git a/docs/llm-hosting/engine-benchmarks-gfx1201.md b/docs/llm-hosting/engine-benchmarks-gfx1201.md
index b6d89fa6f1..126819dff4 100644
--- a/docs/llm-hosting/engine-benchmarks-gfx1201.md
+++ b/docs/llm-hosting/engine-benchmarks-gfx1201.md
@@ -4,10 +4,10 @@ Multi-engine measurement record for Qwen3.6-27B on RDNA4 — vLLM, SGLang and ll
 Measurements are dated and kept as a historical series; the open experiments at the
 bottom are still outstanding.
 
-**Current production engine: SGLang**, `ghcr.io/tanguille/sglang-rdna4`, built by our fork
-of the upstream RDNA4 image rather than a pipeline in this repo — see
-`docs/llm-hosting/sglang-blockers.md` "Current approach" for the build source and the
-upstream cutover plan; outstanding upstream gaps are tracked in the same doc. The
+**Current production engine: vLLM**, serving Qwen 3.8 as `qwen38-27b-vllm` (2026-08-16).
+SGLang and the whole Qwen 3.6 lane were retired with that cutover;
+`docs/llm-hosting/sglang-blockers.md` is kept as the record of why SGLang never
+displaced vLLM here. The
 2026-06-21 round below
 concluded in favour of vLLM and has since been superseded — read its numbers as a
 snapshot of that date, not as current guidance.
diff --git a/docs/llm-hosting/sglang-blockers.md b/docs/llm-hosting/sglang-blockers.md
index 9252ca3a7a..2c78dcccd9 100644
--- a/docs/llm-hosting/sglang-blockers.md
+++ b/docs/llm-hosting/sglang-blockers.md
@@ -2,7 +2,7 @@
 
 Tracking what needs to land upstream before SGLang can replace vLLM in production without depending on the mattbucci RDNA4 fork.
 
-**Current approach:** running `ghcr.io/tanguille/sglang-rdna4:sha-cb7b7605...` (torch 2.11+rocm7.2, Triton 3.6, SGLang v0.5.16), built by our fork's own `build-image.yaml` `publish` job (`Tanguille/2x-R9700-RDNA4-GFX1201-sglang-inference`, `main`), not by a pipeline in this repo — `docker/sglang-rdna4/` and `.github/workflows/build-sglang-rdna4.yaml` are retired. That fork branch carries mattbucci's full patch series plus our HiCache OpenSSL-headers fix (`fix/hicache-openssl-headers`, upstream PR mattbucci#6, still open/unmerged there — this is our own fork build, not the upstream package). `SGLANG_RDNA4_DISABLE_STORE_CACHE=1` is required in the HelmRelease env because LLMKube bypasses the image's entrypoint (see the InferenceService manifest). Once mattbucci#6 merges upstream, re-point at `ghcr.io/mattbucci/sglang-rdna4` directly and drop the fork build. The retired PVC-rebuild recipe is gone with the `sglang` app directory; recover it from git history if ever needed.
+**Historical approach (retired 2026-08-16, superseded by vLLM/Qwen 3.8):** ran `ghcr.io/tanguille/sglang-rdna4:sha-cb7b7605...` (torch 2.11+rocm7.2, Triton 3.6, SGLang v0.5.16), built by our fork's own `build-image.yaml` `publish` job (`Tanguille/2x-R9700-RDNA4-GFX1201-sglang-inference`, `main`), not by a pipeline in this repo — `docker/sglang-rdna4/` and `.github/workflows/build-sglang-rdna4.yaml` are retired. That fork branch carries mattbucci's full patch series plus our HiCache OpenSSL-headers fix (`fix/hicache-openssl-headers`, upstream PR mattbucci#6, still open/unmerged there — this is our own fork build, not the upstream package). `SGLANG_RDNA4_DISABLE_STORE_CACHE=1` was required in the HelmRelease env because LLMKube bypasses the image's entrypoint (see the InferenceService manifest). Once mattbucci#6 merges upstream, re-point at `ghcr.io/mattbucci/sglang-rdna4` directly and drop the fork build. The retired PVC-rebuild recipe is gone with the `sglang` app directory; recover it from git history if ever needed.
 
 **Fork state check (2026-08-01):** the deployed image `sha-cb7b76050cbf` already sits on the fork's v0.5.16 rebase (`689339d`, 2026-07-27), so blocker 6's #31648 (mamba LRU) is live in production — it ships in the v0.5.16 release. Fork patches 086 (AMD Triton `num_kv_splits` 16→64, claimed 2.14× @256K) and 087 (bf16 page-vector attention, claimed +21% @256K) are also in that image; both are model-agnostic (verified via patch content) but were only A/B'd on `coder-reap-25b`, not Qwen3.6-27B. A clean TP=1 re-bench of the deployed image is still outstanding (the 2026-08-01 attempt ran under production load and is recorded as a floor in `engine-benchmarks-gfx1201.md`). `diff cb7b760..origin/main` is docs/eval-only, so a rebuild on current fork main gains no runtime changes today.
 
diff --git a/docs/llm-hosting/vllm-vs-sglang-2026-07.md b/docs/llm-hosting/vllm-vs-sglang-2026-07.md
index 3fd498a9a7..a0c775420e 100644
--- a/docs/llm-hosting/vllm-vs-sglang-2026-07.md
+++ b/docs/llm-hosting/vllm-vs-sglang-2026-07.md
@@ -236,9 +236,10 @@ Both drive the OpenAI `/v1/completions` endpoint, so they run unmodified against
 engine. Port-forward the service and pass the port:
 
 ```sh
-kubectl -n ai port-forward svc/qwen36-27b 30000:30000 &
-python3 concsweep.py 30000
-python3 spectest.py 30000
+kubectl -n ai port-forward svc/qwen38-27b-vllm 8000:8000 &
+python3 concsweep.py 8000
+python3 spectest.py 8000
+# optional 2nd arg overrides the served-model-name (default qwen-3.8)
 ```
 
 **Discard the first run after a pod restart** (see above), and note `concsweep.py` reports
diff --git a/kubernetes/apps/ai/litellm/instance/dashboard/litellm.json b/kubernetes/apps/ai/litellm/instance/dashboard/litellm.json
deleted file mode 100644
index e2e23af1e5..0000000000
--- a/kubernetes/apps/ai/litellm/instance/dashboard/litellm.json
+++ /dev/null
@@ -1,550 +0,0 @@
-{
-  "__inputs": [
-    {
-      "name": "DS_PROMETHEUS",
-      "label": "Prometheus",
-      "type": "datasource",
-      "pluginId": "prometheus",
-      "pluginName": "Prometheus"
-    }
-  ],
-  "description": "LiteLLM proxy metrics. grafana.com 24965 is broken here: job/instance vars use litellm_proxy_total_requests_metric_created, which is not exported (only *_total exists).",
-  "editable": true,
-  "graphTooltip": 1,
-  "links": [
-    {
-      "icon": "external link",
-      "includeVars": true,
-      "keepTime": true,
-      "title": "SGLang",
-      "type": "link",
-      "url": "/d/sglang/sglang"
-    },
-    {
-      "icon": "external link",
-      "includeVars": true,
-      "keepTime": true,
-      "title": "LLMKube",
-      "type": "link",
-      "url": "/d/llmkube-inference/llmkube-inference"
-    }
-  ],
-  "panels": [
-    {
-      "collapsed": false,
-      "gridPos": { "h": 1, "w": 24, "x": 0, "y": 0 },
-      "id": 1,
-      "title": "Overview",
-      "type": "row"
-    },
-    {
-      "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-      "fieldConfig": {
-        "defaults": {
-          "mappings": [
-            {
-              "options": {
-                "0": { "color": "red", "index": 0, "text": "DOWN" },
-                "1": { "color": "green", "index": 1, "text": "UP" }
-              },
-              "type": "value"
-            }
-          ],
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              { "color": "red", "value": null },
-              { "color": "green", "value": 1 }
-            ]
-          }
-        },
-        "overrides": []
-      },
-      "gridPos": { "h": 4, "w": 4, "x": 0, "y": 1 },
-      "id": 2,
-      "options": {
-        "colorMode": "background",
-        "graphMode": "none",
-        "reduceOptions": {
-          "calcs": ["lastNotNull"],
-          "fields": "",
-          "values": false
-        },
-        "textMode": "value"
-      },
-      "targets": [
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "min(up{job=~\"$job\", instance=~\"$instance\"})",
-          "instant": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Scrape Up",
-      "type": "stat"
-    },
-    {
-      "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-      "fieldConfig": {
-        "defaults": { "unit": "reqps", "decimals": 3 },
-        "overrides": []
-      },
-      "gridPos": { "h": 4, "w": 5, "x": 4, "y": 1 },
-      "id": 3,
-      "options": {
-        "colorMode": "value",
-        "graphMode": "area",
-        "reduceOptions": {
-          "calcs": ["lastNotNull"],
-          "fields": "",
-          "values": false
-        },
-        "textMode": "value"
-      },
-      "targets": [
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "sum(rate(litellm_proxy_total_requests_metric_total{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\", status_code=\"200\"}[$__rate_interval]))",
-          "instant": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Success req/s",
-      "type": "stat"
-    },
-    {
-      "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-      "fieldConfig": {
-        "defaults": {
-          "color": { "mode": "thresholds" },
-          "thresholds": {
-            "mode": "absolute",
-            "steps": [
-              { "color": "green", "value": null },
-              { "color": "orange", "value": 0.001 },
-              { "color": "red", "value": 0.01 }
-            ]
-          },
-          "unit": "reqps",
-          "decimals": 3
-        },
-        "overrides": []
-      },
-      "gridPos": { "h": 4, "w": 5, "x": 9, "y": 1 },
-      "id": 4,
-      "options": {
-        "colorMode": "value",
-        "graphMode": "area",
-        "reduceOptions": {
-          "calcs": ["lastNotNull"],
-          "fields": "",
-          "values": false
-        },
-        "textMode": "value"
-      },
-      "targets": [
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "sum(rate(litellm_proxy_total_requests_metric_total{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\", status_code!=\"200\", status_code!=\"None\"}[$__rate_interval]))",
-          "instant": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Error req/s",
-      "type": "stat"
-    },
-    {
-      "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-      "fieldConfig": { "defaults": { "unit": "short" }, "overrides": [] },
-      "gridPos": { "h": 4, "w": 5, "x": 14, "y": 1 },
-      "id": 5,
-      "options": {
-        "colorMode": "value",
-        "graphMode": "area",
-        "reduceOptions": {
-          "calcs": ["lastNotNull"],
-          "fields": "",
-          "values": false
-        },
-        "textMode": "value"
-      },
-      "targets": [
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "sum(litellm_in_flight_requests{job=~\"$job\", instance=~\"$instance\"})",
-          "instant": true,
-          "refId": "A"
-        }
-      ],
-      "title": "In-flight",
-      "type": "stat"
-    },
-    {
-      "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-      "fieldConfig": { "defaults": { "unit": "short" }, "overrides": [] },
-      "gridPos": { "h": 4, "w": 5, "x": 19, "y": 1 },
-      "id": 6,
-      "options": {
-        "colorMode": "value",
-        "graphMode": "none",
-        "reduceOptions": {
-          "calcs": ["lastNotNull"],
-          "fields": "",
-          "values": false
-        },
-        "textMode": "value"
-      },
-      "targets": [
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "sum(litellm_proxy_failed_requests_metric_total{job=~\"$job\", instance=~\"$instance\"})",
-          "instant": true,
-          "refId": "A"
-        }
-      ],
-      "title": "Failed (total)",
-      "type": "stat"
-    },
-    {
-      "collapsed": false,
-      "gridPos": { "h": 1, "w": 24, "x": 0, "y": 5 },
-      "id": 10,
-      "title": "Traffic",
-      "type": "row"
-    },
-    {
-      "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-      "fieldConfig": {
-        "defaults": {
-          "custom": {
-            "drawStyle": "line",
-            "fillOpacity": 10,
-            "lineWidth": 2,
-            "showPoints": "never",
-            "spanNulls": true
-          },
-          "unit": "reqps"
-        },
-        "overrides": []
-      },
-      "gridPos": { "h": 8, "w": 12, "x": 0, "y": 6 },
-      "id": 11,
-      "options": {
-        "legend": {
-          "calcs": ["mean", "max"],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": { "mode": "multi", "sort": "desc" }
-      },
-      "targets": [
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "sum by (requested_model, route, status_code) (rate(litellm_proxy_total_requests_metric_total{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\"}[$__rate_interval]))",
-          "legendFormat": "{{requested_model}} {{route}} {{status_code}}",
-          "refId": "A"
-        }
-      ],
-      "title": "Request rate by model / route",
-      "type": "timeseries"
-    },
-    {
-      "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-      "fieldConfig": {
-        "defaults": {
-          "custom": {
-            "drawStyle": "line",
-            "fillOpacity": 10,
-            "lineWidth": 2,
-            "showPoints": "never",
-            "spanNulls": true
-          },
-          "unit": "ops"
-        },
-        "overrides": []
-      },
-      "gridPos": { "h": 8, "w": 12, "x": 12, "y": 6 },
-      "id": 12,
-      "options": {
-        "legend": {
-          "calcs": ["mean", "max"],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": { "mode": "multi", "sort": "desc" }
-      },
-      "targets": [
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "sum by (requested_model) (rate(litellm_input_tokens_metric_total{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\"}[$__rate_interval]))",
-          "legendFormat": "{{requested_model}} in",
-          "refId": "A"
-        },
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "sum by (requested_model) (rate(litellm_output_tokens_metric_total{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\"}[$__rate_interval]))",
-          "legendFormat": "{{requested_model}} out",
-          "refId": "B"
-        }
-      ],
-      "title": "Token rate",
-      "type": "timeseries"
-    },
-    {
-      "collapsed": false,
-      "gridPos": { "h": 1, "w": 24, "x": 0, "y": 14 },
-      "id": 20,
-      "title": "Latency",
-      "type": "row"
-    },
-    {
-      "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-      "fieldConfig": {
-        "defaults": {
-          "custom": {
-            "drawStyle": "line",
-            "fillOpacity": 10,
-            "lineWidth": 2,
-            "showPoints": "never",
-            "spanNulls": true
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": { "h": 8, "w": 12, "x": 0, "y": 15 },
-      "id": 21,
-      "options": {
-        "legend": {
-          "calcs": ["mean", "max"],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": { "mode": "multi", "sort": "desc" }
-      },
-      "targets": [
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "histogram_quantile(0.5, sum by (requested_model, le) (rate(litellm_llm_api_latency_metric_bucket{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\"}[$__rate_interval])))",
-          "legendFormat": "{{requested_model}} p50",
-          "refId": "A"
-        },
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "histogram_quantile(0.9, sum by (requested_model, le) (rate(litellm_llm_api_latency_metric_bucket{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\"}[$__rate_interval])))",
-          "legendFormat": "{{requested_model}} p90",
-          "refId": "B"
-        },
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "histogram_quantile(0.99, sum by (requested_model, le) (rate(litellm_llm_api_latency_metric_bucket{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\"}[$__rate_interval])))",
-          "legendFormat": "{{requested_model}} p99",
-          "refId": "C"
-        }
-      ],
-      "title": "LLM API latency",
-      "type": "timeseries"
-    },
-    {
-      "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-      "fieldConfig": {
-        "defaults": {
-          "custom": {
-            "drawStyle": "line",
-            "fillOpacity": 10,
-            "lineWidth": 2,
-            "showPoints": "never",
-            "spanNulls": true
-          },
-          "unit": "s"
-        },
-        "overrides": []
-      },
-      "gridPos": { "h": 8, "w": 12, "x": 12, "y": 15 },
-      "id": 22,
-      "options": {
-        "legend": {
-          "calcs": ["mean", "max"],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": { "mode": "multi", "sort": "desc" }
-      },
-      "targets": [
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "histogram_quantile(0.5, sum by (requested_model, le) (rate(litellm_llm_api_time_to_first_token_metric_bucket{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\"}[$__rate_interval])))",
-          "legendFormat": "{{requested_model}} p50",
-          "refId": "A"
-        },
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "histogram_quantile(0.9, sum by (requested_model, le) (rate(litellm_llm_api_time_to_first_token_metric_bucket{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\"}[$__rate_interval])))",
-          "legendFormat": "{{requested_model}} p90",
-          "refId": "B"
-        },
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "histogram_quantile(0.99, sum by (requested_model, le) (rate(litellm_llm_api_time_to_first_token_metric_bucket{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\"}[$__rate_interval])))",
-          "legendFormat": "{{requested_model}} p99",
-          "refId": "C"
-        }
-      ],
-      "title": "Time to first token",
-      "type": "timeseries"
-    },
-    {
-      "collapsed": false,
-      "gridPos": { "h": 1, "w": 24, "x": 0, "y": 23 },
-      "id": 30,
-      "title": "Failures",
-      "type": "row"
-    },
-    {
-      "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-      "fieldConfig": {
-        "defaults": {
-          "custom": {
-            "drawStyle": "line",
-            "fillOpacity": 10,
-            "lineWidth": 2,
-            "showPoints": "never",
-            "spanNulls": true
-          },
-          "unit": "ops"
-        },
-        "overrides": []
-      },
-      "gridPos": { "h": 8, "w": 12, "x": 0, "y": 24 },
-      "id": 31,
-      "options": {
-        "legend": {
-          "calcs": ["mean", "max"],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": { "mode": "multi", "sort": "desc" }
-      },
-      "targets": [
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "sum by (requested_model) (rate(litellm_llm_api_failed_requests_metric_total{job=~\"$job\", instance=~\"$instance\", requested_model=~\"$model\"}[$__rate_interval]))",
-          "legendFormat": "{{requested_model}}",
-          "refId": "A"
-        }
-      ],
-      "title": "LLM API failures / s",
-      "type": "timeseries"
-    },
-    {
-      "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-      "fieldConfig": {
-        "defaults": {
-          "custom": {
-            "drawStyle": "line",
-            "fillOpacity": 10,
-            "lineWidth": 2,
-            "showPoints": "never",
-            "spanNulls": true
-          },
-          "unit": "ops"
-        },
-        "overrides": []
-      },
-      "gridPos": { "h": 8, "w": 12, "x": 12, "y": 24 },
-      "id": 32,
-      "options": {
-        "legend": {
-          "calcs": ["mean", "max"],
-          "displayMode": "table",
-          "placement": "bottom",
-          "showLegend": true
-        },
-        "tooltip": { "mode": "multi", "sort": "desc" }
-      },
-      "targets": [
-        {
-          "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-          "expr": "sum by (litellm_model_name) (rate(litellm_deployment_failure_responses_total{job=~\"$job\", instance=~\"$instance\", litellm_model_name=~\"$model\"}[$__rate_interval]))",
-          "legendFormat": "{{litellm_model_name}}",
-          "refId": "A"
-        }
-      ],
-      "title": "Deployment failures / s",
-      "type": "timeseries"
-    }
-  ],
-  "refresh": "1m",
-  "schemaVersion": 41,
-  "tags": ["ai", "litellm"],
-  "templating": {
-    "list": [
-      {
-        "current": {},
-        "hide": 0,
-        "label": "Datasource",
-        "name": "DS_PROMETHEUS",
-        "options": [],
-        "query": "prometheus",
-        "refresh": 1,
-        "type": "datasource"
-      },
-      {
-        "current": { "selected": true, "text": "litellm", "value": "litellm" },
-        "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-        "definition": "label_values(litellm_proxy_total_requests_metric_total, job)",
-        "includeAll": false,
-        "label": "Job",
-        "name": "job",
-        "query": {
-          "query": "label_values(litellm_proxy_total_requests_metric_total, job)",
-          "refId": "A"
-        },
-        "refresh": 2,
-        "type": "query"
-      },
-      {
-        "allValue": ".*",
-        "current": { "selected": true, "text": "All", "value": "$__all" },
-        "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-        "definition": "label_values(litellm_proxy_total_requests_metric_total{job=~\"$job\"}, instance)",
-        "includeAll": true,
-        "label": "Instance",
-        "multi": true,
-        "name": "instance",
-        "query": {
-          "query": "label_values(litellm_proxy_total_requests_metric_total{job=~\"$job\"}, instance)",
-          "refId": "A"
-        },
-        "refresh": 2,
-        "type": "query"
-      },
-      {
-        "allValue": ".*",
-        "current": { "selected": true, "text": "All", "value": "$__all" },
-        "datasource": { "type": "prometheus", "uid": "${DS_PROMETHEUS}" },
-        "definition": "label_values(litellm_proxy_total_requests_metric_total{job=~\"$job\", instance=~\"$instance\"}, requested_model)",
-        "includeAll": true,
-        "label": "Model",
-        "multi": true,
-        "name": "model",
-        "query": {
-          "query": "label_values(litellm_proxy_total_requests_metric_total{job=~\"$job\", instance=~\"$instance\"}, requested_model)",
-          "refId": "A"
-        },
-        "refresh": 2,
-        "type": "query"
-      }
-    ]
-  },
-  "time": { "from": "now-6h", "to": "now" },
-  "timezone": "browser",
-  "title": "LiteLLM",
-  "uid": "ad8llmv",
-  "version": 1
-}
diff --git a/kubernetes/apps/ai/litellm/instance/grafanadashboard.yaml b/kubernetes/apps/ai/litellm/instance/grafanadashboard.yaml
index cc82526720..c0f5cf5bd1 100644
--- a/kubernetes/apps/ai/litellm/instance/grafanadashboard.yaml
+++ b/kubernetes/apps/ai/litellm/instance/grafanadashboard.yaml
@@ -12,8 +12,8 @@ spec:
   datasources:
     - datasourceName: VictoriaMetrics
       inputName: DS_PROMETHEUS
-  # grafana.com 24965 job/instance vars use litellm_proxy_total_requests_metric_created
-  # (not exported); leaves $job empty and every panel blank.
-  configMapRef:
-    name: litellm-dashboard
-    key: litellm.json
+  # Was a vendored copy: 24965's $job/$instance vars need
+  # litellm_proxy_total_requests_metric_created, unexported until v1.97.0.
+  # No revision -- track latest.
+  grafanaCom:
+    id: 24965
diff --git a/kubernetes/apps/ai/litellm/instance/kustomization.yaml b/kubernetes/apps/ai/litellm/instance/kustomization.yaml
index 7f26aaa074..42277276bc 100644
--- a/kubernetes/apps/ai/litellm/instance/kustomization.yaml
+++ b/kubernetes/apps/ai/litellm/instance/kustomization.yaml
@@ -9,13 +9,3 @@ resources:
   - models.yaml
   - servicemonitor.yaml
   - grafanadashboard.yaml
-
-configMapGenerator:
-  - name: litellm-dashboard
-    files:
-      - dashboard/litellm.json
-    options:
-      annotations:
-        kustomize.toolkit.fluxcd.io/substitute: disabled
-generatorOptions:
-  disableNameSuffixHash: true
diff --git a/kubernetes/apps/ai/litellm/instance/models.yaml b/kubernetes/apps/ai/litellm/instance/models.yaml
index 09fbe9d578..17d2a64262 100644
--- a/kubernetes/apps/ai/litellm/instance/models.yaml
+++ b/kubernetes/apps/ai/litellm/instance/models.yaml
@@ -3,46 +3,55 @@
 apiVersion: litellm.home-operations.com/v1alpha1
 kind: LiteLLMModel
 metadata:
-  name: qwen-3.6
+  name: qwen-3.8
 spec:
-  modelName: qwen-3.6
+  modelName: qwen-3.8
   proxyRef: litellm
   params:
-    model: openai/qwen-3.6
-    apiBase: http://qwen36-27b.ai.svc.cluster.local:30000/v1
-    # Native vLLM/SGLang are unauthenticated internally, but LiteLLM's OpenAI provider
+    # Renamed from qwen-3.6 together with the consumers that hardcode it:
+    # karakeep, agent-pr-review.yaml, proxy.yaml fallbacks. NOT hermes -- its
+    # profiles live on the hermes data PVC, so they are edited by hand and no
+    # rollback of this repo touches them.
+    # Backend swapped to qwen38-27b-vllm 2026-08-16 -- numbers in that file.
+    model: openai/qwen-3.8
+    apiBase: http://qwen38-27b-vllm.ai.svc.cluster.local:8000/v1
+    # Native vLLM is unauthenticated internally, but LiteLLM's OpenAI provider
     # requires a non-empty api_key.
-    apiKey: sk-sglang-noauth
+    apiKey: sk-vllm-noauth
     additional:
       timeout: 1800
       stream_timeout: 1800
       num_retries: 0
   info:
     mode: chat
-    # SGLang's 180000 ctx minus maxOutputTokens.
-    maxInputTokens: 171808
+    # vLLM's maxModelLen (112000) minus maxOutputTokens; covers Hermes' measured
+    # p100 (112K), so routine sessions never truncate. Takes both the
+    # OffloadingConnector and an explicit --kv-cache-memory to reach — see
+    # qwen38-27b-vllm.yaml.
+    maxInputTokens: 103808
     maxOutputTokens: 8192
 ---
 # yaml-language-server: $schema=https://k8s-schemas.home-operations.com/litellm.home-operations.com/litellmmodel_v1alpha1.json
 apiVersion: litellm.home-operations.com/v1alpha1
 kind: LiteLLMModel
 metadata:
-  name: qwen-3.6-fast
+  name: qwen-3.8-fast
 spec:
-  modelName: qwen-3.6-fast
+  modelName: qwen-3.8-fast
   proxyRef: litellm
   params:
-    model: openai/qwen-3.6
-    apiBase: http://qwen36-27b.ai.svc.cluster.local:30000/v1
-    apiKey: sk-sglang-noauth
+    model: openai/qwen-3.8
+    apiBase: http://qwen38-27b-vllm.ai.svc.cluster.local:8000/v1
+    apiKey: sk-vllm-noauth
     additional:
       # Hermes compression is a ~100K+ prefill on this alias; under GPU contention
       # the shared prefill rate drops below what 600s covers (2026-07-07).
       timeout: 900
       stream_timeout: 900
       num_retries: 0
-      # Vendor non-thinking sampling (Qwen3.6 model card) — generation_config ships thinking-mode
-      # values (1.0/0.95/presence 0). Defaults only: client-set params win (v1.91.0 router merge).
+      # Non-thinking instruct-mode sampling — matches Qwen3.8's model card
+      # exactly (temperature/top_p/presence_penalty), same as Qwen3.6's.
+      # Defaults only: client-set params win (v1.91.0 router merge).
       temperature: 0.7
       top_p: 0.8
       presence_penalty: 1.5
@@ -51,7 +60,8 @@ spec:
           enable_thinking: false
   info:
     mode: chat
-    maxInputTokens: 171808
+    # 112000 - 8192, as above.
+    maxInputTokens: 103808
     maxOutputTokens: 8192
 ---
 # yaml-language-server: $schema=https://k8s-schemas.home-operations.com/litellm.home-operations.com/litellmmodel_v1alpha1.json
@@ -68,10 +78,10 @@ spec:
     model: openai/auto/smart
     apiBase: http://omniroute.ai.svc.cluster.local:20128/v1
     # REQUIRE_API_KEY=false internally; LiteLLM's OpenAI provider still needs
-    # a non-empty api_key (same convention as the qwen-3.6 sk-sglang-noauth).
+    # a non-empty api_key (same convention as qwen-3.8's sk-vllm-noauth).
     apiKey: sk-omniroute-noauth
     additional:
-      # Bounded so the qwen-3.6-fast fallback fires inside the pr-reviewer-action's
+      # Bounded so the qwen-3.8-fast fallback fires inside the pr-reviewer-action's
       # 600s curl timeout instead of the action giving up and retrying blind
       # (observed 2026-08-03: 60min run, fallback never triggered). Timeouts
       # aren't in the proxy's retry_policy and num_retries is 0, so one attempt
diff --git a/kubernetes/apps/ai/litellm/instance/proxy.yaml b/kubernetes/apps/ai/litellm/instance/proxy.yaml
index 9b026537a8..923d618afd 100644
--- a/kubernetes/apps/ai/litellm/instance/proxy.yaml
+++ b/kubernetes/apps/ai/litellm/instance/proxy.yaml
@@ -49,17 +49,17 @@ spec:
     drop_params: true
     num_retries: 0
   routerSettings:
-    # Per-error override of num_retries: 0 — a dropped SGLang connection costs a
+    # Per-error override of num_retries: 0 — a dropped backend connection costs a
     # whole Hermes cron run. Timeouts stay at 0: re-running a ~100K prefill that
     # already burned 900s only deepens the contention that caused it.
     retry_policy:
       InternalServerErrorRetries: 2
     # omniroute's free providers have no SLA and can fail/time out without
     # warning; fall back to the self-hosted (slower but always-available)
-    # qwen-3.6-fast alias so a flaky free provider doesn't take out the PR
+    # qwen-3.8-fast alias so a flaky free provider doesn't take out the PR
     # review action.
     fallbacks:
-      - omniroute: ["qwen-3.6-fast"]
+      - omniroute: ["qwen-3.8-fast"]
   route:
     hostnames:
       - "litellm.${SECRET_DOMAIN}"
diff --git a/kubernetes/apps/ai/llmkube/app/helmrelease.yaml b/kubernetes/apps/ai/llmkube/app/helmrelease.yaml
index 560a2554a4..01436b8751 100644
--- a/kubernetes/apps/ai/llmkube/app/helmrelease.yaml
+++ b/kubernetes/apps/ai/llmkube/app/helmrelease.yaml
@@ -41,10 +41,11 @@ spec:
       dashboards:
         enabled: true
         namespace: observability
-        # The rest render blank here: amd-gpu-observability queries amdgpu_*
-        # (we run drm-exporter), llmkube-quota and model-router need features we
-        # don't deploy, llmkube-slo needs pyrra burn-rate rules.
-        only: [llmkube-inference, sglang-dashboard]
+        # No `only`: ship everything and let the operator's default auto mode
+        # gate the runtime dashboards on a live InferenceService. Auto covers
+        # sglang/vllm/llamacpp only, so amd-gpu-observability, llmkube-quota,
+        # llmkube-slo and model-router ship unconditionally and render blank
+        # (no pyrra, and we export drm_* not amdgpu_*).
         operator:
           enabled: true
           folder: ai
@@ -58,7 +59,7 @@ spec:
     # Shared RWX cache on CephFS: the controller and the iGPU InferenceServices
     # land on different nodes, so an RWO volume multi-attach-errors. Access is
     # write-once on download then read-once at load, so CephFS latency never
-    # touches inference. qwen36-27b brings its own node-local modelCache.
+    # touches inference. qwen38-27b-vllm brings its own node-local modelCache.
     # The model-cache-prep init chowns /models as root, so CephFS (fsGroupPolicy:
     # None) works too.
     modelCache:
diff --git a/kubernetes/apps/ai/llmkube/models/kustomization.yaml b/kubernetes/apps/ai/llmkube/models/kustomization.yaml
index 75802d4772..778ea9dc8d 100644
--- a/kubernetes/apps/ai/llmkube/models/kustomization.yaml
+++ b/kubernetes/apps/ai/llmkube/models/kustomization.yaml
@@ -3,9 +3,8 @@
 apiVersion: kustomize.config.k8s.io/v1beta1
 kind: Kustomization
 resources:
-  - qwen36-27b-sglang.yaml
+  - qwen38-27b-vllm.yaml
   - qwen3-embedding.yaml
   - qwen35-2b.yaml
   - vmcp-embedding.yaml
-  - qwen36-27b-vllm.yaml
   - muse-glimmer-30b.yaml
diff --git a/kubernetes/apps/ai/llmkube/models/muse-glimmer-30b.yaml b/kubernetes/apps/ai/llmkube/models/muse-glimmer-30b.yaml
index 63c98f65c6..39fe901265 100644
--- a/kubernetes/apps/ai/llmkube/models/muse-glimmer-30b.yaml
+++ b/kubernetes/apps/ai/llmkube/models/muse-glimmer-30b.yaml
@@ -48,7 +48,7 @@ spec:
   modelRef: muse-glimmer-30b
   modelCache:
     claimName: muse-glimmer-30b-model-cache
-  # Config is evaluated and kept; scale to 1 only with qwen36-27b-vllm at 0.
+  # Config is evaluated and kept; scale to 1 only with qwen38-27b-vllm at 0.
   replicas: 0
   # Needs >= b10361 for muse_glimmer.
   image: ghcr.io/ggml-org/llama.cpp:server-rocm-b10454@sha256:799571945f574c640d1151633f49c202cd778bf5a227df6ef314a1f08623f5d3
diff --git a/kubernetes/apps/ai/llmkube/models/qwen3-embedding.yaml b/kubernetes/apps/ai/llmkube/models/qwen3-embedding.yaml
index 7f741b2055..3d9b74da58 100644
--- a/kubernetes/apps/ai/llmkube/models/qwen3-embedding.yaml
+++ b/kubernetes/apps/ai/llmkube/models/qwen3-embedding.yaml
@@ -59,7 +59,8 @@ spec:
     - "0"
 
   # Pin to the iGPU node; without this the scheduler could place it on the dGPU
-  # node (which also advertises squat.ai/dri) and steal a dGPU slot from sglang.
+  # node (which also advertises squat.ai/dri) and steal a dGPU slot from the
+  # 27B serving pod.
   nodeSelector:
     amd.com/igpu: "true"
 
diff --git a/kubernetes/apps/ai/llmkube/models/qwen36-27b-sglang.yaml b/kubernetes/apps/ai/llmkube/models/qwen36-27b-sglang.yaml
deleted file mode 100644
index e95540e9b0..0000000000
--- a/kubernetes/apps/ai/llmkube/models/qwen36-27b-sglang.yaml
+++ /dev/null
@@ -1,336 +0,0 @@
----
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: qwen36-27b-model-cache
-  namespace: ai
-spec:
-  accessModes:
-    - ReadWriteOnce
-  storageClassName: openebs-hostpath
-  resources:
-    requests:
-      storage: 50Gi
----
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: qwen36-27b-triton-cache
-  namespace: ai
-spec:
-  accessModes:
-    - ReadWriteOnce
-  storageClassName: openebs-hostpath
-  resources:
-    requests:
-      storage: 10Gi
----
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: qwen36-27b-hicache
-  namespace: ai
-spec:
-  accessModes:
-    - ReadWriteOnce
-  storageClassName: openebs-hostpath
-  resources:
-    # hostpath enforces no quota, so the real bound is the evictor's MAX_SIZE below.
-    requests:
-      storage: 64Gi
----
-# Prunes HiCache generations superseded by an image bump (why they are versioned at all:
-# see SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR below). This CRD has no initContainers field
-# to automate it, so it was manual, and one missed cleanup orphaned 64G on 2026-07-30 --
-# against 60.3G free on control-1's sparse TrueNAS zvol, where filling the pool stops the
-# VM. Tradeoff: rolling back >1 day after a bump finds that cache pruned (cold TTFT).
-apiVersion: batch/v1
-kind: CronJob
-metadata:
-  name: qwen36-27b-hicache-prune
-  namespace: ai
-spec:
-  schedule: "30 4 * * *"
-  concurrencyPolicy: Forbid
-  successfulJobsHistoryLimit: 1
-  failedJobsHistoryLimit: 3
-  jobTemplate:
-    spec:
-      backoffLimit: 2
-      template:
-        spec:
-          restartPolicy: OnFailure
-          securityContext:
-            # Superseded generations were written by root-era pods (pre-4a4b06f).
-            runAsUser: 0
-          containers:
-            - name: prune
-              image: ghcr.io/home-operations/busybox:1.38.0@sha256:7e2c04dd50ede647bf4a7a4c8dbd629dd4971cd139b9b88fb22bfc3c7a6c13df
-              securityContext:
-                allowPrivilegeEscalation: false
-                readOnlyRootFilesystem: true
-                capabilities:
-                  drop: ["ALL"]
-              command:
-                - /bin/sh
-                - -c
-                - |
-                  set -eu
-                  cd /hicache
-                  # Newest mtime = the live generation; a serving pod keeps promoting into
-                  # it. Skip it and anything written in the last day, so a draining pod
-                  # keeps its cache through a rollout.
-                  newest="$(ls -1dt -- */ 2>/dev/null | head -1)"
-                  for d in */; do
-                    [ -d "$d" ] || continue
-                    [ "$d" = "$newest" ] && continue
-                    [ -n "$(find "$d" -maxdepth 0 -mmin +1440)" ] || continue
-                    echo "pruning superseded generation: $d"
-                    rm -rf -- "$d"
-                  done
-                  # Names/counts only: du here would stat the whole store (682,508 files on
-                  # 2026-07-30) on top of the walk rm already does, just to log a number.
-                  echo "remaining: $(ls -1d -- */ 2>/dev/null | wc -l) generation(s)"
-              volumeMounts:
-                - name: hicache
-                  mountPath: /hicache
-              resources:
-                requests:
-                  cpu: 10m
-                  memory: 64Mi
-                # Memory only: busybox rm streams, so this just bounds the pathological
-                # case. No cpu limit -- the job is iowait-bound, where CFS quota cannot
-                # throttle, and if entries are cache-hot it would stretch the run instead.
-                limits:
-                  memory: 128Mi
-          volumes:
-            - name: hicache
-              persistentVolumeClaim:
-                claimName: qwen36-27b-hicache
----
-# yaml-language-server: $schema=https://k8s-schemas.home-operations.com/inference.llmkube.dev/model_v1alpha1.json
-apiVersion: inference.llmkube.dev/v1alpha1
-kind: Model
-metadata:
-  name: qwen36-27b-awq
-spec:
-  source: hf://mattbucci/Qwen3.6-27B-AWQ@f541031dfc1bf5da3f198b52b9bf44cbd5ceee49
-  format: safetensors
-  refreshPolicy: OnChange
-  files:
-    - model.safetensors
-    - model-vision.safetensors
-    - model.safetensors.index.json
-    - config.json
-    - generation_config.json
-    - chat_template.jinja
-    - processor_config.json
-    - tokenizer.json
-    - tokenizer_config.json
-  hardware:
-    accelerator: rocm
-    gpu:
-      enabled: true
-      count: 1
-      vendor: amd
-      runtime: rocm
-      resourceName: squat.ai/dri
----
-# yaml-language-server: $schema=https://k8s-schemas.home-operations.com/inference.llmkube.dev/inferenceservice_v1alpha1.json
-apiVersion: inference.llmkube.dev/v1alpha1
-kind: InferenceService
-metadata:
-  name: qwen36-27b
-spec:
-  modelRef: qwen36-27b-awq
-  runtime: sglang
-  # Owns the R9700: more context, faster decode and working vision than vLLM at 0.
-  replicas: 1
-  # Operator default binds :: (IPv6-only on this IPv4 cluster); override to v4.
-  bindAddress: 0.0.0.0
-  image: ghcr.io/tanguille/sglang-rdna4:sha-cb7b76050cbf8dda9e5a78c07b24df1a783caeb3@sha256:f961fc274eaad4af065097e0921b88ae5a97a478baa0e384e6c5ee440c355b29
-  containerPort: 30000
-  endpoint:
-    # v0.9.8 maps endpoint.port to both Service port and targetPort; use the
-    # runtime port so generated traffic reaches SGLang directly.
-    port: 30000
-  nodeSelector:
-    kubernetes.io/hostname: control-1
-  podSecurityContext:
-    # The image's own USER 10001:10001 owns /home/sglang/.cache 0700 (flashinfer's JIT
-    # workspace lives there). The operator drops ALL capabilities on the container, so
-    # runAsUser: 0 no longer gets CAP_DAC_OVERRIDE to force its way past that ownership --
-    # root became just another uid. Match the image's own user instead of overriding it.
-    runAsUser: 10001
-    runAsGroup: 10001
-    # The triton-cache and hicache PVCs were written by every prior root-based pod, so
-    # they're root-owned on disk. fsGroup makes kubelet recursively chown volume mounts
-    # to this group at mount time -- the same fix as runAsGroup, but for PVC contents
-    # instead of the image's own baked files. OnRootMismatch: hicache already holds 64G;
-    # skip the recursive chown walk on every future restart once the root matches.
-    fsGroup: 10001
-    fsGroupChangePolicy: OnRootMismatch
-    supplementalGroups:
-      - 44
-      - 226
-    seccompProfile:
-      type: Unconfined
-  modelCache:
-    claimName: qwen36-27b-model-cache
-  sglangConfig:
-    contextLength: 180000
-    # HARD CEILING: the R9700 is shared with Jellyfin transcoding, raising this fails
-    # media workloads. 0.95 grows the pool 183,240 -> 261,019 tokens but OOMs on a
-    # 13.8K-token prefill (2026-07-25).
-    memFractionStatic: 0.875
-    # No maxRunningRequests: the mamba clamp below always wins over a requested value
-    # (resolve_max_num_reqs), so setting one here would only misstate the real ceiling.
-    chunkedPrefillSize: 4096
-    # No tensorParallelSize (1 is the default) or quantization: config.json carries
-    # quant_method awq and dtype bfloat16, both auto-detected; awq_marlin, the one override
-    # that could diverge, hard-returns False off CUDA (awq.py:388). Raising TP also means
-    # re-adding the flags dropped as TP=1 no-ops under extraArgs. kvCacheDtype does NOT
-    # auto-detect -- it buys the pool.
-    kvCacheDtype: fp8_e4m3
-    reasoningParser: qwen3
-  resources:
-    cpu: "2"
-    hostMemory: 24Gi
-  env:
-    # No *_VISIBLE_DEVICES/GPU_DEVICE_ORDINAL: the device plugin hands over one render node
-    # (card0 + renderD128) and torch sees one device, so "0" was the identity mapping x4.
-    # $HOME/.cache/flashinfer/<ver>/ is root-owned in the image (a build-time
-    # `import torch, sglang, sgl_kernel` verification step triggers flashinfer's JIT
-    # logger init as root, before the runtime user exists) -- uid 10001 can enter it
-    # (0755) but not write the JIT log, crashing at import. Redirect off $HOME entirely.
-    - name: FLASHINFER_WORKSPACE_BASE
-      value: /tmp
-    - name: TRITON_CACHE_DIR
-      value: /cache/sglang/triton
-    # Required at TP=1: LLMKube runs launch_server directly, bypassing the image's
-    # entrypoint.sh/gpu-selection.sh, the only place this is normally set. Without it the
-    # JIT store_cache kernel crashes on first request (kvcache.cuh:204).
-    - name: SGLANG_RDNA4_DISABLE_STORE_CACHE
-      value: "1"
-    # Page names encode neither the weights revision nor the engine, so this suffix MUST be
-    # bumped with the image digest above or the AWQ revision, or stale KV pages get served.
-    # SGLang's own per-model cache key (config_suffix, HiCacheFile.__init__) only covers
-    # --served-model-name/TP, which stays "qwen-3.6" across bumps -- it doesn't catch this.
-    # Its evictor only manages its own configured path, so the old suffix's directory is
-    # orphaned on every bump; the qwen36-27b-hicache-prune CronJob above reclaims it.
-    - name: SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR
-      value: /hicache/sglang-v0.5.16_awq-f541031d_fork-cb7b760
-    # Unset defaults to unbounded. 64Gi made L3 write-only: it evicted every entry before
-    # it was read back. 128Gi didn't fix it either -- the store still sat at ~99% full
-    # continuously (2026-08-04 measurement), so it was still evicting under write
-    # pressure just as before, only less visibly. 192Gi gives real headroom: control-1's
-    # /var had 268.7Gi free and SSD_Pool (backing the zvol) had 364Gi free at time of
-    # writing. Deliberately above the PVC's nominal 64Gi -- hostpath enforces no quota
-    # and openebs-hostpath cannot expand, so this env var is the real bound.
-    - name: SGLANG_HICACHE_FILE_BACKEND_MAX_SIZE
-      value: 192Gi
-    # MAX_SIZE is blind to the other tenants of control-1's shared root, so hold a floor
-    # above kubelet's nodefs eviction threshold: refuse writes rather than fill the disk.
-    # 160Gi, not 100Gi: rook-ceph's mon-c store shares this 500G filesystem and MON_DISK_LOW
-    # fires below 30% avail (150Gi), which a 100Gi floor guarantees (2026-08-06).
-    - name: SGLANG_HICACHE_FILE_BACKEND_MIN_FREE_SPACE
-      value: 160Gi
-    # Not the eviction trigger despite the log's wording: eviction fires at MAX_SIZE and
-    # this is the low-water mark it drains to. Higher = warmer cache, more frequent passes.
-    - name: SGLANG_HICACHE_FILE_BACKEND_EVICTION_RATIO
-      value: "0.95"
-    # Without this a lookup miss walks the whole store with os.scandir. The startup scan it
-    # adds is one os.listdir, not a per-file stat. Off by default.
-    - name: SGLANG_HICACHE_FILE_BACKEND_ENABLE_METADATA_CACHE
-      value: "1"
-    # The 5s default expires every entry the startup scan just inserted; -1 never expires.
-    - name: SGLANG_HICACHE_FILE_BACKEND_METADATA_TTL
-      value: "-1"
-  extraVolumes:
-    - name: triton-cache
-      persistentVolumeClaim:
-        claimName: qwen36-27b-triton-cache
-    - name: hicache
-      persistentVolumeClaim:
-        claimName: qwen36-27b-hicache
-    - name: dshm
-      emptyDir:
-        medium: Memory
-        sizeLimit: 8Gi
-  extraVolumeMounts:
-    - name: triton-cache
-      mountPath: /cache
-    - name: hicache
-      mountPath: /hicache
-    - name: dshm
-      mountPath: /dev/shm
-  probeOverrides:
-    # SGLang's health endpoint is suitable for startup; keep liveness inert so
-    # long model loads or transient GPU work cannot trigger a restart.
-    startup:
-      httpGet:
-        path: /health
-        port: 30000
-      initialDelaySeconds: 60
-      periodSeconds: 15
-      timeoutSeconds: 5
-      failureThreshold: 120
-    liveness:
-      exec:
-        command:
-          - "true"
-      # Inert probe; hourly period avoids a pointless CRI exec every 10s.
-      periodSeconds: 3600
-    readiness:
-      tcpSocket:
-        port: 30000
-      periodSeconds: 30
-      timeoutSeconds: 5
-      failureThreshold: 6
-  extraArgs:
-    # Dropped as no-ops at TP=1 on this checkpoint: --dtype bfloat16 (auto-resolves),
-    # --trust-remote-code (no auto_map, no .py files), --pre-warm-nccl (gated on tp/pp/ep>1,
-    # bootstrap.py:109), --disable-custom-all-reduce (never built at world_size 1, :438).
-    - --num-continuous-decode-steps
-    - "16"
-    - --watchdog-timeout
-    - "600"
-    - --attention-backend
-    - triton
-    - --chat-template
-    - /opt/rdna4-inference/scripts/qwen3.6_devrole_chat_template.jinja
-    - --tool-call-parser
-    - qwen3_coder
-    - --disable-overlap-schedule
-    - --cuda-graph-backend-decode=disabled
-    - --cuda-graph-backend-prefill=disabled
-    # Mamba slots and KV tokens share one byte budget (unified_memory_pool.py), so unused
-    # slots are KV tokens forgone: 4.46G/60 slots = 74.3MB each vs 5.60G/183,240 KV tokens.
-    # 14d utilization (3643 5m samples): slots hit 32 in only 2 samples (0.055%), never higher.
-    # 32 matches the observed peak with zero margin, freeing 16 slots (~1.2G) to the KV pool.
-    - --max-mamba-cache-size
-    - "32"
-    - --served-model-name
-    - qwen-3.6
-    - --mamba-ssm-dtype
-    - bfloat16
-    - --max-queued-requests
-    - "32"
-    - --weight-loader-drop-cache-after-load
-    - --enable-hierarchical-cache
-    - --hicache-io-backend
-    - direct
-    - --hicache-ratio
-    - "1.5"
-    - --model-loader-extra-config
-    - '{"num_threads":2}'
-    - --enable-mixed-chunk
-    # Without this SGLang returns prompt_tokens_details:None and hermes logs
-    # cache_read_tokens=0 regardless of the real hit rate.
-    - --enable-cache-report
-    # Promote only on hit_count >= 2. page_size is pinned at 1 by mamba no_buffer, so
-    # write_through is ~1 file per prompt token, which halved per-request decode.
-    - --hicache-write-policy
-    - write_through_selective
-    - --hicache-storage-backend
-    - file
diff --git a/kubernetes/apps/ai/llmkube/models/qwen36-27b-vllm.yaml b/kubernetes/apps/ai/llmkube/models/qwen36-27b-vllm.yaml
deleted file mode 100644
index 8408191eb9..0000000000
--- a/kubernetes/apps/ai/llmkube/models/qwen36-27b-vllm.yaml
+++ /dev/null
@@ -1,264 +0,0 @@
----
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: qwen36-27b-vllm-cache
-  namespace: ai
-spec:
-  accessModes:
-    - ReadWriteOnce
-  storageClassName: openebs-hostpath
-  resources:
-    requests:
-      storage: 50Gi
----
-# Persist torch, Triton, and Inductor compile caches across restarts.
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: qwen36-27b-vllm-compile-cache
-  namespace: ai
-spec:
-  accessModes:
-    - ReadWriteOnce
-  storageClassName: openebs-hostpath
-  resources:
-    requests:
-      storage: 10Gi
----
-# L3 tier for the fs secondary tier -- the vLLM analogue of SGLang's HiCache file backend.
-apiVersion: v1
-kind: PersistentVolumeClaim
-metadata:
-  name: qwen36-27b-vllm-kv-offload
-  namespace: ai
-spec:
-  accessModes:
-    - ReadWriteOnce
-  storageClassName: openebs-hostpath
-  resources:
-    # hostpath enforces no quota, so the real bound is the free-space guard below.
-    requests:
-      storage: 128Gi
----
-# vLLM's fs tier has no evictor at v0.27.1, and control-1's root fs is shared with
-# rook-ceph's mon-c store. Enforces the 160Gi floor SGLang's file backend enforces itself.
-apiVersion: batch/v1
-kind: CronJob
-metadata:
-  name: qwen36-27b-vllm-kv-offload-guard
-  namespace: ai
-spec:
-  # Daily: SGLang's store took weeks to reach 147G.
-  schedule: "0 4 * * *"
-  concurrencyPolicy: Forbid
-  successfulJobsHistoryLimit: 1
-  failedJobsHistoryLimit: 3
-  jobTemplate:
-    spec:
-      backoffLimit: 2
-      template:
-        spec:
-          restartPolicy: OnFailure
-          securityContext:
-            # The serving pod runs runAsUser 0, so the store is root-owned on disk.
-            runAsUser: 0
-          containers:
-            - name: guard
-              image: ghcr.io/home-operations/busybox:1.38.0@sha256:7e2c04dd50ede647bf4a7a4c8dbd629dd4971cd139b9b88fb22bfc3c7a6c13df
-              securityContext:
-                allowPrivilegeEscalation: false
-                readOnlyRootFilesystem: true
-                capabilities:
-                  drop: ["ALL"]
-              command:
-                - /bin/sh
-                - -c
-                - |
-                  # $$ throughout: Flux postBuild envsubst substitutes bare $VAR.
-                  set -eu
-                  floor=167772160
-                  # df, not du: the store reaches hundreds of thousands of files.
-                  avail="$$(df -kP /kvoffload | awk 'NR==2{print $$4}')"
-                  echo "avail: $${avail}K (floor: $${floor}K)"
-                  if [ "$$avail" -lt "$$floor" ]; then
-                    # ponytail: wipe-all, not LRU -- no per-entry recency on disk to sort by.
-                    echo "below floor, wiping the L3 store"
-                    rm -rf -- /kvoffload/*
-                  fi
-              volumeMounts:
-                - name: kv-offload
-                  mountPath: /kvoffload
-              resources:
-                requests:
-                  cpu: 10m
-                  memory: 64Mi
-                limits:
-                  memory: 128Mi
-          volumes:
-            - name: kv-offload
-              persistentVolumeClaim:
-                claimName: qwen36-27b-vllm-kv-offload
----
-# yaml-language-server: $schema=https://k8s-schemas.home-operations.com/inference.llmkube.dev/model_v1alpha1.json
-apiVersion: inference.llmkube.dev/v1alpha1
-kind: Model
-metadata:
-  name: qwen36-27b-vllm
-spec:
-  # Pin the model revision for reproducible staging. QuantTrio is the boot-validated
-  # variant; the prior cyankiwi INT4 pin was never booted.
-  source: hf://QuantTrio/Qwen3.6-27B-AWQ@9b507bdc9afafb87b7898700cc2a591aa6639461
-  format: safetensors
-  quantization: AWQ
-  files:
-    - model-00001-of-00008.safetensors
-    - model-00002-of-00008.safetensors
-    - model-00003-of-00008.safetensors
-    - model-00004-of-00008.safetensors
-    - model-00005-of-00008.safetensors
-    - model-00006-of-00008.safetensors
-    - model-00007-of-00008.safetensors
-    - model-00008-of-00008.safetensors
-    - model.safetensors.index.json
-    - config.json
-    - configuration.json
-    - generation_config.json
-    - tokenizer.json
-    - tokenizer_config.json
-    - chat_template.jinja
-    - merges.txt
-    - vocab.json
-    - preprocessor_config.json
-    - video_preprocessor_config.json
-  refreshPolicy: OnChange
-  hardware:
-    accelerator: rocm
-    memoryBudget: 34Gi
-    gpu:
-      enabled: true
-      count: 1
-      vendor: amd
-      runtime: rocm
-      resourceName: squat.ai/dri
-      memory: 32Gi
----
-# yaml-language-server: $schema=https://k8s-schemas.home-operations.com/inference.llmkube.dev/inferenceservice_v1alpha1.json
-apiVersion: inference.llmkube.dev/v1alpha1
-kind: InferenceService
-metadata:
-  name: qwen36-27b-vllm
-spec:
-  modelRef: qwen36-27b-vllm
-  runtime: vllm
-  # Operator default binds :: (IPv6-only on this IPv4 cluster); override to v4.
-  bindAddress: 0.0.0.0
-  modelCache:
-    claimName: qwen36-27b-vllm-cache
-  # v0.26.0 measured identical to the nightly on KV pool and throughput (2026-07-25).
-  # The nightly self-reports a stale 0.23.1rc1 base tag; tags are not version-ordered.
-  image: vllm/vllm-openai-rocm:v0.27.1@sha256:bb44b39aea26798cce43030a98bf48efd0322ca7147367db86e38b96bd80f0e7
-  # Parked: 3.9 tok/s decode on real traffic, and vision needs the pool it cannot spare.
-  replicas: 0
-  # Above 16 the 125,041-token pool oversubscribes: TG 48 -> 35, TTFT 11s -> 85s.
-  parallelSlots: 16
-  vllmConfig:
-    # Fits the measured pool below; SGLang's 180K needs ~6.55 GiB.
-    maxModelLen: 112000
-    kvCacheDtype: fp8_e4m3
-    # Off, mamba leaves 'align' mode and the offloading connector asserts.
-    enablePrefixCaching: true
-    maxNumBatchedTokens: 4096
-    # HARD CEILING: shared with Jellyfin transcoding, never raise.
-    gpuMemoryUtilization: 0.875
-  env:
-    - name: VLLM_ROCM_USE_AITER
-      value: "0"
-    - name: HIP_VISIBLE_DEVICES
-      value: "0"
-    - name: ROCR_VISIBLE_DEVICES
-      value: "0"
-    - name: VLLM_CACHE_ROOT
-      value: /cache/vllm
-    - name: TRITON_CACHE_DIR
-      value: /cache/triton
-    - name: TORCHINDUCTOR_CACHE_DIR
-      value: /cache/inductor
-    # Pins the fs tier's block-hash seed; unset it is random per process and restarts lose it.
-    - name: PYTHONHASHSEED
-      value: "0"
-  extraVolumes:
-    - name: compile-cache
-      persistentVolumeClaim:
-        claimName: qwen36-27b-vllm-compile-cache
-    - name: kv-offload
-      persistentVolumeClaim:
-        claimName: qwen36-27b-vllm-kv-offload
-    # The CPU offload tier mmaps 8Gi here; the 64M default EFAULTs on pre-fault.
-    - name: dshm
-      emptyDir:
-        medium: Memory
-        sizeLimit: 10Gi
-  extraVolumeMounts:
-    - name: compile-cache
-      mountPath: /cache
-    - name: kv-offload
-      mountPath: /kvoffload
-    - name: dshm
-      mountPath: /dev/shm
-  extraArgs:
-    - --served-model-name
-    - qwen-3.6
-    # Zeroes every modality limit, so the vision tower loads on the meta device.
-    - --language-model-only
-    # Use bfloat16 Mamba cache dtype to reduce memory for the long context window.
-    - --mamba-ssm-cache-dtype
-    - bfloat16
-    - --trust-remote-code
-    # Without these the gateway gets thinking blocks and tool calls as raw text.
-    - --reasoning-parser
-    - qwen3
-    - --enable-auto-tool-choice
-    - --tool-call-parser
-    - qwen3_coder
-    # vLLM's own "to fit into requested memory" figure at 0.875 (measured 2026-08-12).
-    - --kv-cache-memory-bytes
-    - "4691494912"
-    # Hierarchical KV cache: CPU tier = SGLang's --hicache-ratio (9.01 GB measured), fs tier
-    # = its file backend. No block_size -- Qwen3.6's hybrid groups trip a divisibility assert.
-    - --kv-transfer-config
-    - >-
-      {"kv_connector":"OffloadingConnector","kv_role":"kv_both","kv_connector_extra_config":{"spec_name":"TieringOffloadingSpec","cpu_bytes_to_use":8589934592,"secondary_tiers":[{"type":"fs","root_dir":"/kvoffload","locality":"LOCAL"}]}}
-  probeOverrides:
-    startup:
-      httpGet:
-        path: /health
-        port: 8000
-      initialDelaySeconds: 60
-      periodSeconds: 15
-      failureThreshold: 240
-    # Inert by design: long model loads and transient GPU work must not restart the pod.
-    liveness:
-      exec:
-        command: ["true"]
-      periodSeconds: 3600
-    readiness:
-      tcpSocket:
-        port: 8000
-      periodSeconds: 30
-      timeoutSeconds: 5
-      failureThreshold: 6
-  nodeSelector:
-    amd.com/gpu: "true"
-  podSecurityContext:
-    runAsUser: 0
-    runAsGroup: 0
-    supplementalGroups: [44, 226]
-    seccompProfile:
-      type: Unconfined
-  resources:
-    cpu: "2"
-    # 24Gi base + the 8 GiB the CPU offload tier pins.
-    memory: 32Gi
-  endpoint:
-    port: 8000
diff --git a/kubernetes/apps/ai/llmkube/models/qwen38-27b-vllm.yaml b/kubernetes/apps/ai/llmkube/models/qwen38-27b-vllm.yaml
new file mode 100644
index 0000000000..be674401ed
--- /dev/null
+++ b/kubernetes/apps/ai/llmkube/models/qwen38-27b-vllm.yaml
@@ -0,0 +1,311 @@
+---
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: qwen38-27b-vllm-model-cache
+spec:
+  accessModes: [ReadWriteOnce]
+  storageClassName: openebs-hostpath
+  resources:
+    requests:
+      storage: 50Gi
+---
+# Persist torch, Triton, and Inductor compile caches across restarts.
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: qwen38-27b-vllm-compile-cache
+spec:
+  accessModes: [ReadWriteOnce]
+  storageClassName: openebs-hostpath
+  resources:
+    requests:
+      storage: 10Gi
+---
+# L3 tier for the fs secondary tier -- the vLLM analogue of SGLang's HiCache
+# file backend. Ported from qwen36-27b-vllm.yaml: proven pattern, same
+# hardware, same GPU-memory ceiling.
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: qwen38-27b-vllm-kv-offload
+spec:
+  accessModes: [ReadWriteOnce]
+  storageClassName: openebs-hostpath
+  resources:
+    # hostpath enforces no quota, so the real bound is the free-space guard below.
+    requests:
+      storage: 128Gi
+---
+# vLLM's fs tier has no evictor at v0.27.1, and this node's root fs is shared with
+# rook-ceph's mon store, so the tier needs an external one.
+apiVersion: batch/v1
+kind: CronJob
+metadata:
+  name: qwen38-27b-vllm-kv-offload-guard
+spec:
+  # Hourly, not daily: a daily window let a busy day cross the floor and sit
+  # there for hours. Safe to run this often only because eviction is now
+  # incremental (see the script) rather than a full wipe.
+  # ponytail: periodic sweep, not a hard cap -- move to a quota-enforcing
+  # storage class if the floor is ever actually breached.
+  schedule: "0 * * * *"
+  concurrencyPolicy: Forbid
+  successfulJobsHistoryLimit: 1
+  failedJobsHistoryLimit: 3
+  jobTemplate:
+    spec:
+      backoffLimit: 2
+      template:
+        spec:
+          restartPolicy: OnFailure
+          securityContext:
+            # The serving pod runs runAsUser 0, so the store is root-owned on disk.
+            runAsUser: 0
+          containers:
+            - name: guard
+              image: ghcr.io/home-operations/busybox:1.38.0@sha256:7e2c04dd50ede647bf4a7a4c8dbd629dd4971cd139b9b88fb22bfc3c7a6c13df
+              securityContext:
+                allowPrivilegeEscalation: false
+                readOnlyRootFilesystem: true
+                capabilities:
+                  drop: ["ALL"]
+              command:
+                - /bin/sh
+                - -c
+                - |
+                  # $$ throughout: Flux postBuild envsubst substitutes bare $VAR.
+                  set -eu
+                  # 80Gi, not the 160Gi this started as: unreachable on a 500G
+                  # disk with ~370G held elsewhere, so the guard wiped the whole
+                  # tier every run. The prefix-hit rate needs the store to survive.
+                  floor=83886080
+                  # df, not du: the store reaches hundreds of thousands of files.
+                  avail="$$(df -kP /kvoffload | awk 'NR==2{print $$4}')"
+                  echo "avail: $${avail}K (floor: $${floor}K)"
+                  if [ "$$avail" -ge "$$floor" ]; then
+                    echo "above floor, nothing to do"
+                    exit 0
+                  fi
+                  # Oldest-first, not a wipe: blocks are write-once so mtime is
+                  # real recency. Walk down only until back above the floor.
+                  for age in 30 14 7 3 1 0; do
+                    find /kvoffload -type f -mtime +$$age -delete 2>/dev/null || true
+                    avail="$$(df -kP /kvoffload | awk 'NR==2{print $$4}')"
+                    echo "evicted >$${age}d, avail now $${avail}K"
+                    if [ "$$avail" -ge "$$floor" ]; then
+                      exit 0
+                    fi
+                  done
+                  # Still short with only <24h blocks left: the shortfall is
+                  # elsewhere on the shared fs. Fail loudly -- a guard that
+                  # silently misses its floor reads healthy until the node fills.
+                  echo "floor unreachable: $${avail}K < $${floor}K, only <24h blocks left" >&2
+                  exit 1
+              volumeMounts:
+                - name: kv-offload
+                  mountPath: /kvoffload
+              resources:
+                requests:
+                  cpu: 10m
+                  memory: 64Mi
+                limits:
+                  memory: 128Mi
+          volumes:
+            - name: kv-offload
+              persistentVolumeClaim:
+                claimName: qwen38-27b-vllm-kv-offload
+---
+# yaml-language-server: $schema=https://k8s-schemas.home-operations.com/inference.llmkube.dev/model_v1alpha1.json
+apiVersion: inference.llmkube.dev/v1alpha1
+kind: Model
+metadata:
+  name: qwen38-27b-vllm
+spec:
+  source: hf://cyankiwi/Qwen3.8-27B-AWQ-INT4@63768c10df38c0395e12ef49edac1bd539eaeeea
+  format: safetensors
+  quantization: compressed-tensors
+  refreshPolicy: OnChange
+  files:
+    - model-00001-of-00005.safetensors
+    - model-00002-of-00005.safetensors
+    - model-00003-of-00005.safetensors
+    - model-00004-of-00005.safetensors
+    - model-00005-of-00005.safetensors
+    - model.safetensors.index.json
+    - config.json
+    - generation_config.json
+    - tokenizer.json
+    - tokenizer_config.json
+    - chat_template.jinja
+    - merges.txt
+    - vocab.json
+    - preprocessor_config.json
+    - video_preprocessor_config.json
+  hardware:
+    accelerator: rocm
+    memoryBudget: 28Gi
+    gpu:
+      enabled: true
+      count: 1
+      vendor: amd
+      runtime: rocm
+      resourceName: squat.ai/dri
+      memory: 32Gi
+---
+# yaml-language-server: $schema=https://k8s-schemas.home-operations.com/inference.llmkube.dev/inferenceservice_v1alpha1.json
+# concsweep.py, warm, agg tok/s vs the SGLang qwen-3.6 baseline (conc 1/8/16):
+# 14.88/53.93/100.82 vs 14.96/36.24/34.49. Bench warm only -- cold reads ~12.
+apiVersion: inference.llmkube.dev/v1alpha1
+kind: InferenceService
+metadata:
+  name: qwen38-27b-vllm
+spec:
+  modelRef: qwen38-27b-vllm
+  runtime: vllm
+  replicas: 1
+  # Matches the proven qwen36 production concurrency ceiling; operator
+  # translates this to --max-num-seqs for the vllm runtime.
+  parallelSlots: 16
+  bindAddress: 0.0.0.0
+  image: vllm/vllm-openai-rocm:v0.27.1@sha256:bb44b39aea26798cce43030a98bf48efd0322ca7147367db86e38b96bd80f0e7
+  modelCache:
+    claimName: qwen38-27b-vllm-model-cache
+  vllmConfig:
+    # Hermes hard-requires >=64K context (refuses to start below it); its
+    # measured percentiles are p90 94K, p95 106K, p100 112K, so this covers
+    # p100 and routine sessions never truncate. Reachable only with the
+    # explicit --kv-cache-memory below.
+    maxModelLen: 112000
+    kvCacheDtype: fp8_e4m3
+    # Confirmed working on this hybrid Mamba/GDN model (~27% hit rate under
+    # real traffic on the sibling qwen36 config).
+    enablePrefixCaching: true
+    maxNumBatchedTokens: 4096
+    # INERT while --kv-cache-memory is set below (vLLM skips profiling and logs
+    # that it ignores this). Kept because dropping that flag hands sizing back
+    # here, and 0.875 is the hard ceiling: the ~4.6 GB free at this value is
+    # Jellyfin's transcode reservation, not slack, and nothing enforces it
+    # (docs/llm-hosting/vllm-vs-sglang-2026-07.md).
+    gpuMemoryUtilization: 0.875
+  env:
+    - name: VLLM_ROCM_USE_AITER
+      value: "0"
+    - name: HIP_VISIBLE_DEVICES
+      value: "0"
+    - name: ROCR_VISIBLE_DEVICES
+      value: "0"
+    - name: VLLM_CACHE_ROOT
+      value: /cache/vllm
+    - name: TRITON_CACHE_DIR
+      value: /cache/triton
+    - name: TORCHINDUCTOR_CACHE_DIR
+      value: /cache/inductor
+    # Pins the fs tier's block-hash seed; unset it is random per process and restarts lose it.
+    - name: PYTHONHASHSEED
+      value: "0"
+  extraVolumes:
+    - name: compile-cache
+      persistentVolumeClaim:
+        claimName: qwen38-27b-vllm-compile-cache
+    - name: kv-offload
+      persistentVolumeClaim:
+        claimName: qwen38-27b-vllm-kv-offload
+    # The CPU offload tier mmaps 8Gi here; the 64M default EFAULTs on pre-fault.
+    - name: dshm
+      emptyDir:
+        medium: Memory
+        sizeLimit: 10Gi
+  extraVolumeMounts:
+    - name: compile-cache
+      mountPath: /cache
+    - name: kv-offload
+      mountPath: /kvoffload
+    - name: dshm
+      mountPath: /dev/shm
+  extraArgs:
+    - --served-model-name
+    - qwen-3.8
+    - --trust-remote-code
+    - --reasoning-parser
+    - qwen3
+    - --enable-auto-tool-choice
+    - --tool-call-parser
+    - qwen3_coder
+    # Qwen3.5-family GDN hybrids OOM during CUDA-graph capture at vLLM's
+    # default 512-size ceiling regardless of gpuMemoryUtilization/
+    # enablePrefixCaching (vLLM #39010 on gfx1201). Cap capture sizes to the
+    # production concurrency ceiling (16) instead of disabling graphs
+    # outright, to keep decode throughput on par with 3.6.
+    # Also fixes a real crash, not just perf: hybrid Mamba/GDN cache only had
+    # 71 blocks available at this budget, and vLLM's default max-num-seqs
+    # (256) overshoots it — spec.parallelSlots above must stay <= that.
+    - --compilation-config
+    - '{"cudagraph_capture_sizes": [1, 2, 4, 8, 16]}'
+    # Keep prior-turn reasoning in context; matches production qwen36 config.
+    - --default-chat-template-kwargs
+    - '{"preserve_thinking": true}'
+    # MTP speculative decoding is deliberately OMITTED: measured on this
+    # hardware, MTP + tool-calling grammar wedges to ~0.2 tok/s (100x
+    # regression) under concurrent tool traffic. Re-test that wedge before
+    # re-adding.
+    # Covers the conv-state and SSM pools both (--mamba-ssm-cache-dtype
+    # defaults to following this), and bfloat16 is the smallest vLLM accepts.
+    - --mamba-cache-dtype
+    - bfloat16
+    # Skips profiling, so this -- not gpuMemoryUtilization -- sizes KV. The
+    # profiler was conservative: 3.22 GiB, which forced the old 98K ceiling.
+    # 5 GiB is the most that still clears Jellyfin's ~4.6 GB: 5.46 GB free,
+    # 156,493 tokens, 1.40x at 112K. 6 GiB gives 1.68x but only 3.41 GB free,
+    # under the reservation. Re-measure free VRAM before raising.
+    - --kv-cache-memory
+    - "5368709120"
+    # Hierarchical KV cache: CPU tier (matches SGLang's --hicache-ratio sizing
+    # philosophy) + fs tier (SGLang's file backend analogue). Ported from
+    # qwen36-27b-vllm.yaml, WITHOUT that config's --language-model-only —
+    # Hermes' auxiliary.vision block actively uses this model's vision tower,
+    # so it must stay loaded even though that costs some context headroom.
+    - --kv-transfer-config
+    - >-
+      {"kv_connector":"OffloadingConnector","kv_role":"kv_both","kv_connector_extra_config":{"spec_name":"TieringOffloadingSpec","cpu_bytes_to_use":8589934592,"secondary_tiers":[{"type":"fs","root_dir":"/kvoffload","locality":"LOCAL"}]}}
+  nodeSelector:
+    amd.com/gpu: "true"
+  podSecurityContext:
+    runAsUser: 0
+    runAsGroup: 0
+    supplementalGroups: [44, 226]
+    seccompProfile:
+      type: Unconfined
+  resources:
+    cpu: "2"
+    # Measured steady state is 13.3Gi (8Gi of it the shm offload tier, which is
+    # tmpfs and so charged here). Kept at 32Gi for the model-load spike rather
+    # than trimmed to the steady state -- an OOMKill mid-load throws away the
+    # compile cache this node takes ~40min to rebuild.
+    memory: 32Gi
+  endpoint:
+    port: 8000
+  probeOverrides:
+    startup:
+      httpGet:
+        path: /health
+        port: 8000
+      initialDelaySeconds: 60
+      periodSeconds: 15
+      failureThreshold: 240
+    # Startup owns the load/compile window, so this only ever sees a serving
+    # engine. Slack on purpose: a restart costs ~2min, so only a 10-minute
+    # outage (20 x 30s) counts as death.
+    liveness:
+      httpGet:
+        path: /health
+        port: 8000
+      periodSeconds: 30
+      timeoutSeconds: 10
+      failureThreshold: 20
+    readiness:
+      tcpSocket:
+        port: 8000
+      periodSeconds: 30
+      timeoutSeconds: 5
+      failureThreshold: 6
diff --git a/kubernetes/apps/ai/omniroute/app/helmrelease.yaml b/kubernetes/apps/ai/omniroute/app/helmrelease.yaml
index 5ee3afd4ea..320da5f240 100644
--- a/kubernetes/apps/ai/omniroute/app/helmrelease.yaml
+++ b/kubernetes/apps/ai/omniroute/app/helmrelease.yaml
@@ -31,7 +31,7 @@ spec:
               DATA_DIR: /app/data
               PORT: &port 20128
               # Only litellm's /v1/* proxy calls skip auth (unauthenticated-backend
-              # + non-empty-apiKey convention, same as the qwen-3.6 SGLang model);
+              # + non-empty-apiKey convention, same as the qwen-3.8 vLLM model);
               # the dashboard at the internal route below still requires
               # INITIAL_PASSWORD login.
               REQUIRE_API_KEY: "false"
diff --git a/kubernetes/apps/default/karakeep/app/helmrelease.yaml b/kubernetes/apps/default/karakeep/app/helmrelease.yaml
index 463e0ff80d..4d517d90cc 100644
--- a/kubernetes/apps/default/karakeep/app/helmrelease.yaml
+++ b/kubernetes/apps/default/karakeep/app/helmrelease.yaml
@@ -55,7 +55,7 @@ spec:
               # its LiteLLMModel is auto-registered (litellm-operator autoRegister)
               # under the InferenceService's name, not a hand-written alias.
               INFERENCE_TEXT_MODEL: qwen35-2b
-              INFERENCE_IMAGE_MODEL: qwen-3.6-fast
+              INFERENCE_IMAGE_MODEL: qwen-3.8-fast
               INFERENCE_NUM_WORKERS: "2"
               INFERENCE_OUTPUT_SCHEMA: "plain"
               INFERENCE_ENABLE_AUTO_SUMMARIZATION: "true"