diff --git a/pmoves/config/profiles/laptop-4090.yaml b/pmoves/config/profiles/laptop-4090.yaml
index 43b323090b..6589b184f9 100644
--- a/pmoves/config/profiles/laptop-4090.yaml
+++ b/pmoves/config/profiles/laptop-4090.yaml
@@ -61,6 +61,8 @@ coding_stacks:
     tz_function: coding_alibaba
 
 # VRAM budget (16GB total, 2GB system reserve = 14GB usable)
+# Installed variants: qwen3-embedding:8b (4.7GB) and qwen3-vl:8b (6.1GB).
+# Profile updated 2026-05-10 to match actual inventory.
 vram_budget:
   total_mb: 16384
   system_reserve_mb: 2048
@@ -68,22 +70,22 @@ vram_budget:
   max_concurrent_models: 2
   idle_timeout_seconds: 180
   always_resident:
-    - model: qwen3-embedding:4b
-      vram_mb: 3000
+    - model: qwen3-embedding:8b        # 8b installed (original design specified :4b)
+      vram_mb: 4700
       reason: "HiRAG, extract-worker, cipher-memory need constant embeddings"
   co_residency_groups:
     coding_focus:
-      models: [qwen3-coder:30b, qwen3-embedding:4b]
-      total_mb: 9500
+      models: [qwen3-coder:30b, qwen3-embedding:8b]
+      total_mb: 11000                  # 30b offloads excess layers to 64GB RAM
     general_focus:
-      models: [qwen3.5:9b, qwen3-embedding:4b]
-      total_mb: 9600
+      models: [qwen3.5:9b, qwen3-embedding:8b]
+      total_mb: 11000
     research_focus:
-      models: [lfm2:24b, qwen3-embedding:4b]
-      total_mb: 14000                # Tight — may need embedding eviction (usable: 14336)
+      models: [lfm2:24b, qwen3-embedding:8b]
+      total_mb: 14336                  # Tight — may need embedding eviction
     multimodal_minimal:
-      models: [qwen3.5:4b, qwen3-embedding:4b, qwen3-vl:4b]
-      total_mb: 9000
+      models: [qwen3.5:4b, qwen3-embedding:8b, qwen3-vl:8b]
+      total_mb: 13800                  # 3GB + 4.7GB + 6.1GB — fits 14.3GB
 
 mcp:
   - docker
@@ -143,6 +145,15 @@ cross_node:
         - nats:4222
         - minio:9000
       access_via: [tailscale]
+    pmoves-spark:
+      role: fleet_inference
+      tailscale_host: pmoves-spark
+      hardware: "GB10 Grace-Blackwell 128GB unified — ARM64"
+      services:
+        - ollama:11434                 # Gemma 4 31B FP16, Nemotron-Super-49B
+        - nim:8200                     # NVIDIA NIM (optional)
+      access_via: [tailscale, nats]
+      notes: "Primary 120B+ inference node. Route lfm2:24b and qwen3-coder-480b here at full precision."
 
   channels:
     nats:
@@ -174,7 +185,7 @@ nats_announce:
     vram: 16gb
     cuda: sm_89
     coding: [qwen3-coder-30b, qwen3.5-9b]
-    embeddings: qwen3-embed-4b
+    embeddings: qwen3-embed-8b
     stacks: [claude, codex, glm, minimax, alibaba]
     claws: 4
     geometry_bus: true
diff --git a/pmoves/configs/tac_trees/node-4090-laptop.tac.yaml b/pmoves/configs/tac_trees/node-4090-laptop.tac.yaml
index fe409e3e8e..3253cf034d 100644
--- a/pmoves/configs/tac_trees/node-4090-laptop.tac.yaml
+++ b/pmoves/configs/tac_trees/node-4090-laptop.tac.yaml
@@ -148,8 +148,8 @@ root:
     # Phase 4: Tailscale Mesh Connectivity
     # =========================================================================
     - id: n4090.mesh
-      task: "Tailscale mesh connectivity to 5090 and Z890"
-      context: "All 3 nodes must be reachable for full fleet operation"
+      task: "Tailscale mesh connectivity to 5090, Z890, and SPARK"
+      context: "All fleet nodes must be reachable for full operation"
       agent_hint: 4090-claude
       children:
         - id: n4090.mesh.status
@@ -160,6 +160,18 @@ root:
             expect: "pmoves-laptop online at pmoves-laptop.ts"
           agent_hint: 4090-claude
 
+        - id: n4090.mesh.ssh-enabled
+          task: "Tailscale SSH enabled on 4090 laptop"
+          action:
+            type: shell
+            command: "tailscale debug features 2>/dev/null | grep -i ssh || tailscale status --json | python -c \"import json,sys; d=json.load(sys.stdin); print('ssh-enabled' if d.get('Self',{}).get('TailscaleSSHEnabled') else 'DISABLED')\""
+            expect: "ssh-enabled (or SSH feature listed in debug features output)"
+          context: >
+            Tailscale SSH allows SPARK and other fleet nodes to SSH into
+            pmoves-laptop without Windows password. Auth via Tailscale identity.
+            Enable with: tailscale set --ssh (operator runs once).
+          agent_hint: 4090-claude
+
         - id: n4090.mesh.reach-5090
           task: "Can reach 5090 (POWERFULMOVES)"
           action:
@@ -177,6 +189,17 @@ root:
             expect: "Pong from pmoves-z890.ts via direct connection"
           agent_hint: 4090-claude
 
+        - id: n4090.mesh.reach-spark
+          task: "Can reach SPARK (pmoves-spark)"
+          action:
+            type: shell
+            command: "tailscale ping pmoves-spark --timeout 5s --c 1"
+            expect: "Pong from pmoves-spark via direct connection"
+          context: >
+            DGX Spark GB10 Grace-Blackwell 128GB — fleet inference backbone.
+            Primary 120B+ model host. Must be tagged tag:gpu in Tailscale admin.
+          agent_hint: 4090-claude
+
         - id: n4090.mesh.z890-services
           task: "Z890 key services reachable"
           action:
@@ -189,6 +212,62 @@ root:
             expect: "All Z890 services accessible via Tailscale IP"
           agent_hint: 4090-claude
 
+    # =========================================================================
+    # Phase 4b: Ollama Model Health
+    # =========================================================================
+    - id: n4090.ollama
+      task: "Ollama model inventory matches laptop-4090.yaml profile"
+      context: "Profile expects: qwen3-embedding:8b, qwen3-vl:8b, qwen3-coder:30b, qwen3.5:9b, qwen3.5:4b, lfm2:24b"
+      agent_hint: 4090-claude
+      children:
+        - id: n4090.ollama.running
+          task: "Ollama service running on :11434"
+          action:
+            type: shell
+            command: "curl -s http://localhost:11434/api/tags | python -c \"import json,sys; names=[m['name'] for m in json.load(sys.stdin)['models']]; required={'qwen3-embedding:8b','qwen3-vl:8b','qwen3-coder:30b','qwen3.5:9b','qwen3.5:4b','lfm2:24b'}; missing=required-set(names); print('ok' if not missing else 'MISSING:'+','.join(sorted(missing)))\""
+            expect: "ok (all 6 profile-expected models present)"
+          agent_hint: 4090-claude
+
+        - id: n4090.ollama.embedding
+          task: "qwen3-embedding:8b resident (always-resident)"
+          action:
+            type: shell
+            command: "curl -s http://localhost:11434/api/tags | python -c \"import json,sys; names=[m['name'] for m in json.load(sys.stdin)['models']]; print('ok' if 'qwen3-embedding:8b' in names else 'MISSING')\""
+            expect: "ok"
+          context: >
+            Always-resident embedding model (4.7GB VRAM).
+            If missing: ollama pull qwen3-embedding:8b
+          agent_hint: 4090-claude
+
+        - id: n4090.ollama.coder
+          task: "qwen3-coder:30b available (coding fallback)"
+          action:
+            type: shell
+            command: "curl -s http://localhost:11434/api/tags | python -c \"import json,sys; names=[m['name'] for m in json.load(sys.stdin)['models']]; print('ok' if 'qwen3-coder:30b' in names else 'MISSING')\""
+            expect: "ok"
+          context: >
+            Primary local coding fallback. ~18GB on disk — partial CPU offload.
+            If missing: ollama pull qwen3-coder:30b (~18GB, 10+ min)
+          agent_hint: 4090-claude
+
+        - id: n4090.ollama.stale-check
+          task: "No pre-2025 stale models on disk"
+          action:
+            type: manual
+            instruction: >
+              Run: ollama list
+              Verify NONE of these appear: gemma2, gemma3, llama3.2,
+              deepseek-r1, deepseek-coder-v2, phi4, qwen2.5, nomic-embed-text,
+              embeddinggemma, llama3.2-vision, jina-embeddings-v2-base-code, functiongemma
+            expect: "Only profile-expected models present"
+          context: >
+            Stale models consume ~67GB disk. Cleanup:
+            ollama rm nomic-embed-text embeddinggemma gemma3:latest gemma3:12b
+            gemma2:latest llama3.2:latest llama3.2-vision qwen2.5:7b qwen2.5vl:7b
+            deepseek-r1:latest deepseek-coder-v2:latest phi4:latest
+            jina-embeddings-v2-base-code functiongemma
+          agent_hint: 4090-claude
+
     # =========================================================================
     # Phase 5: NATS Announcement
     # =========================================================================