NVIDIA · wenxind-nvidia · Jul 30, 2026 · Aug 11, 2026
diff --git a/DEPENDENCIES.md b/DEPENDENCIES.md
@@ -111,7 +111,8 @@ xr-ai-models  (agent-sdk/xr-ai-models/)
     seam: reasoning-field aliasing (nano_v3 →
     `reasoning`, nemotron_v3 → `reasoning_content`), `chat_template_kwargs`
     plumbing for `enable_thinking` / `thinking_budget`, and built-in presets
-    for the in-tree services. Future backends (LiteLLM, vendor SDKs)
+    for the in-tree services, including distinct Cosmos3 Nano Reasoner and
+    Cosmos-Reason1 VLM profiles. Future backends (LiteLLM, vendor SDKs)
     plug in as new `kind`s in `_factory.py::make_*` without touching the
     protocols or callers.  Workers depend on this instead of rolling their
     own httpx wrappers. Profiles may separate adapter, endpoint, and deployment
@@ -199,7 +200,9 @@ xr-ai-vllm  (utils/xr-ai-vllm/)
     `--stop` flow.  Besides `serve` / `stop_persistent_servers`, exposes the
     shared wrapper helpers `resolve_model_cache`, `load_config`, `setup_hf_env`,
     and `gpu_compute_major` (all stdlib-only; pyyaml is imported function-locally
-    inside `load_config` so the `--stop` path stays dependency-free).
+    inside `load_config` so the `--stop` path stays dependency-free). Persistent
+    Docker containers and pip processes carry launch fingerprints; stale
+    repository-owned runtimes are replaced before reuse.
 
 xr-ai-vad  (utils/xr-ai-vad/)
     └── numpy >=1.24
@@ -371,12 +374,15 @@ xr-ai-tests  (tests/)
     `video-mcp-server` rather than redeclared here.
 
 vlm-server  (services/vlm-server/)
-    └── vllm >=0.12.0
+    └── vllm >=0.23.0
     └── pyyaml >=6.0
     └── hf-transfer >=0.1.4
     └── xr-ai-logging  [editable: ../../utils/xr-ai-logging]
     └── xr-ai-vllm     [editable: ../../utils/xr-ai-vllm]
-    Model: nvidia/Cosmos-Reason1-7B (Qwen2.5-VL architecture, vLLM).
+    Default model: the text-output Reasoner from nvidia/Cosmos3-Nano. Standard
+    vLLM serving loads only the Reasoner; the Generator's vLLM-Omni path is
+    outside this service. nvidia/Cosmos-Reason1-7B remains configurable with
+    the cosmos_vlm client preset.
     Wrapper Popens `vllm serve` so the launcher's killpg() does not reach
     vLLM — model survives stack restarts (see docs/changelog.md 2026-05-05).
     vllm_backend: pip|docker — pip path uses the wrapper's vllm; docker path
@@ -470,7 +476,7 @@ piper-tts-server  (services/piper-tts/)
 
 | Server | Package | Command | Default port | Model | Backend |
 |---|---|---|---|---|---|
-| `services/vlm-server/` | `vlm-server` | `vlm_server` | 8100 | Cosmos-Reason1-7B | vLLM (pip or docker) |
+| `services/vlm-server/` | `vlm-server` | `vlm_server` | 8100 | Cosmos3 Nano Reasoner | vLLM (pip or docker) |
 | `services/stt-server/` | `stt-server` | `stt_server` | 8103 | parakeet-tdt-0.6b-v3 | NeMo ASR in-process |
 | `services/magpie-tts/` | `magpie-tts-server` | `magpie_tts_server` | 8104 | magpie_tts_multilingual_357m | NeMo TTS in-process |
 | `services/piper-tts/` | `piper-tts-server` | `piper_tts_server` | 8105 | rhasspy/piper-voices (ONNX) | piper-tts in-process |
@@ -558,9 +564,10 @@ remain provided by the shared voice runtime. The sample has no direct
 Worker calls stt-server (8103), vlm-server (8100), and piper-tts-server
 (8105) over HTTP via `xr-ai-models` SDK — no model weights loaded
 in-process. The `models_config` key selects a structured deployment profile:
-`models.local.json` manages the default services, `models.hosted.json` uses an
-external NVIDIA NIM VLM, and `models.omni.json` reuses Nemotron-Omni on port
-8108. These profiles separate adapter behavior, endpoint readiness and
+`models.local.json` manages the default Cosmos3 Nano Reasoner service,
+`models.hosted.json` uses the hosted Cosmos3 Nano Reasoner NIM, and
+`models.omni.json` reuses Nemotron-Omni on port 8108. These profiles separate
+adapter behavior, endpoint readiness and
 credentials, and launcher ownership. Voice-gate knobs are configured via
 `yaml/voice_gate.yaml`.
 
@@ -628,7 +635,7 @@ updated in the same commit**.
 | `vllm_backend` / `vllm_image` YAML keys | `services/{vlm-server,embedding-server,llama-nemotron-llm,nemotron3-nano-llm,nemotron-omni-llm}/<server>.yaml`, every per-profile copy in `agent-samples/`, `docs/ai-services.md` |
 | Container name used by a vllm wrapper | `_CONTAINER_NAME` in the wrapper's `__main__.py`, `stop_persistent_servers` names in `agent-samples/model-servers/main.py` |
 | vlm-server model class or supported architectures | `services/vlm-server/vlm_server.yaml` comments |
-| vlm-server YAML config keys (`model`, `model_cache`, …) | `services/vlm-server/vlm_server.yaml`, `agent-samples/simple-vlm-example/vlm_server.yaml` |
+| vlm-server YAML config keys (`model`, `model_cache`, `async_scheduling`, …) | `services/vlm-server/vlm_server.yaml`, every per-profile copy in `agent-samples/` |
 | cloudxr-runtime YAML config keys | `agent-samples/xr-render-demo/yaml/cloudxr_runtime.yaml`, `docs/adding-cloudxr.md` |
 | `utils/xr-ai-launcher/xr_ai_launcher/_cloudxr_env.py` API | xr-render-scene + oxr-mcp + `services/cloudxr-runtime/cloudxr_runtime/__main__.py` imports, `agent-samples/xr-render-demo/main.py` (native-profile gate), `docs/adding-cloudxr.md`, `docs/xr-render-demo.md` (client-type section) |
 | scene service YAML config keys | `agent-samples/xr-render-demo/scene/scene_service.yaml`, orchestrator process declaration, `docs/xr-render-demo.md` |

diff --git a/README.md b/README.md
@@ -161,7 +161,7 @@ uv run model_servers
 ```
 
 GPU profiles are auto-detected (`dual_48G_ada` / `spark` / `96G_blackwell`).
-On first run each model downloads from HuggingFace (~50 GB total; can take
+On first run each model downloads from HuggingFace (tens of GB; can take
 tens of minutes).  On subsequent runs the containers restart in under a minute.
 
 The default `--vlm-llm-stack` starts Nemotron-3 Nano (8107), Cosmos (8100),
@@ -176,7 +176,7 @@ they cannot be stopped, avoiding GPU overcommit.
 uv run model_servers --omni-stack
 ```
 
-`HF_TOKEN` is required by default: without it the ~50 GB first-run download
+`HF_TOKEN` is required by default: without it the large first-run download
 can stall indefinitely.  See [`docs/credentials.md`](docs/credentials.md)
 for how to set it, or pass `--allow-anonymous` to run without one.
 
@@ -201,7 +201,8 @@ MCP client is involved. See the
 [sample README](agent-samples/simple-vlm-example/README.md) for the worker
 layout and configuration boundaries.
 
-Uses `nvidia/Cosmos-Reason1-7B` (NVIDIA Open Model License + Apache 2.0).
+Uses the text-output Reasoner from `nvidia/Cosmos3-Nano` by default. Standard
+vLLM serving loads only the Reasoner, not the separate Generator pipeline.
 
 There are two ways to run it:
 
@@ -214,7 +215,7 @@ uv sync
 uv run simple_vlm_example
 ```
 
-On the very first run weights download from HuggingFace (~23 GB; can take
+On the very first run weights download from HuggingFace (tens of GB; can take
 several minutes).  `HF_TOKEN` is required by default; pass
 `--allow-anonymous` to run without one (see
 [`docs/credentials.md`](docs/credentials.md)).
@@ -265,6 +266,9 @@ after a moment, and you hear the reply through your speakers.
 **Local model** — override the model weights or GPU settings by editing
 `vlm_server.yaml` in the sample directory.
 
+To use Cosmos-Reason1 instead, set `model: nvidia/Cosmos-Reason1-7B` in that
+file and select `cosmos_vlm` as the VLM adapter preset in `models.local.json`.
+
 **Remote model** — copy `yaml/models.hosted.json`, point its VLM endpoint at
 your server, and select it in the worker config:
 

diff --git a/agent-mcp-servers/vlm-mcp/vlm_mcp_server.yaml b/agent-mcp-servers/vlm-mcp/vlm_mcp_server.yaml
@@ -19,19 +19,16 @@ host:                  0.0.0.0
 port:                  8240
 
 # Model configuration — identifies the VLM backend by logical name "vlm".
-# The cosmos_vlm preset targets services/vlm-server/ (Cosmos-Reason1-7B
-# via vLLM). Only base_url needs to be set here; the preset fills in the rest.
+# The cosmos3_nano_reasoner preset targets services/vlm-server/. Only base_url
+# needs to be set here; the preset fills in the rest.
 models:
   vlm:
-    kind:     preset:cosmos_vlm
+    kind:     preset:cosmos3_nano_reasoner
     base_url: http://localhost:8100
 
 # httpx timeout applied to each call to vlm-server.
 vlm_request_timeout_s: 60.0
 
-# Whether to let the VLM produce chain-of-thought reasoning before answering.
-# false (default) — suppresses <think> generation entirely; faster responses,
-#                   better for real-time captioning and voice loops.
-# true             — enables full reasoning; more accurate on complex queries
-#                   but adds several seconds of latency per request.
+# Forward an enable_thinking chat-template flag for backends that support it.
+# Cosmos3 Nano Reasoner does not use this Qwen-specific switch.
 enable_thinking: false
diff --git a/agent-mcp-servers/vlm-mcp/vlm_mcp_server/__main__.py b/agent-mcp-servers/vlm-mcp/vlm_mcp_server/__main__.py
@@ -35,7 +35,7 @@
     port:                 8240
     models:
       vlm:
-        kind:     preset:cosmos_vlm
+        kind:     preset:cosmos3_nano_reasoner
         base_url: http://localhost:8100
     vlm_request_timeout_s: 60.0
     enable_thinking: false
@@ -100,8 +100,8 @@ def _make_vlm_from_cfg(cfg: dict[str, Any]) -> tuple[VLMService, float]:
         if nested:
             vlm_entry["endpoint"] = target
 
-        # The cosmos_vlm preset defaults enable_thinking to False; an explicit
-        # top-level true must reach the wire by overriding default_extras.
+        # An explicit top-level true must reach backends that support the
+        # enable_thinking chat-template flag.
         if enable_thinking:
             target = dict(vlm_entry.get("adapter") or {}) if nested else vlm_entry
             extras = dict(target.get("default_extras") or {})
@@ -119,7 +119,7 @@ def _make_vlm_from_cfg(cfg: dict[str, Any]) -> tuple[VLMService, float]:
     elif vlm_server:
         logger.warning(
             "vlm_mcp_server.yaml: 'vlm_server' key is deprecated — "
-            "migrate to a 'models:' block with kind: preset:cosmos_vlm"
+            "migrate to a 'models:' block with an explicit VLM preset"
         )
         chat_template_kwargs: dict[str, Any] = {"enable_thinking": enable_thinking}
         config = load_models_config_from_dict({

diff --git a/agent-samples/model-servers/main.py b/agent-samples/model-servers/main.py
@@ -12,7 +12,7 @@
   default / --vlm-llm-stack
     stt        — nvidia/parakeet-tdt-0.6b-v3        port 8103  (NeMo ASR)
     agent-llm  — NVIDIA-Nemotron-3-Nano-30B-A3B     port 8107  (vLLM)
-    vlm        — nvidia/Cosmos-Reason1-7B           port 8100  (vLLM)
+    vlm        — nvidia/Cosmos3-Nano Reasoner       port 8100  (vLLM)
     embedding  — nvidia/llama-nemotron-embed-1b-v2  port 8109  (vLLM)
 
   --omni-stack

diff --git a/agent-samples/model-servers/yaml/96G_blackwell/vlm_server.yaml b/agent-samples/model-servers/yaml/96G_blackwell/vlm_server.yaml
@@ -5,22 +5,22 @@
 #
 # vlm is SECOND in load order: agent-llm → vlm.
 # free_at_startup ≈ 96 − 34 (agent-llm) − 2 (STT) ≈ 60 GiB.
-# 0.20 × 96 = 19.2 GiB cap: ~16 GiB Cosmos-7B weights + ~3 GiB KV.
+# 0.20 × 96 = 19.2 GiB cap: ~16 GiB Cosmos3 Reasoner weights + ~3 GiB KV.
 
-model: nvidia/Cosmos-Reason1-7B
+model: nvidia/Cosmos3-Nano
 hf_token: ""
 port: 8100
 host: "0.0.0.0"
 model_cache: ../../../../models
 gpu_memory_utilization: 0.20
 max_num_seqs: 4
 enforce_eager: true
+async_scheduling: true
 max_videos_per_prompt: 0
-# 4 fits within the default max_model_len=8192 (each Qwen2.5-VL image is
-# ~1.5k tokens). Raise both together if multi-image turns get larger.
+# Raise max_model_len too if multi-image turns outgrow the context window.
 max_images_per_prompt: 4
 
 # vLLM runtime: pip (this wrapper's venv) or docker (NGC nvcr.io/nvidia/vllm).
 # See services/vlm-server/vlm_server.yaml for the full toggle docs.
 vllm_backend: docker
-vllm_image:   nvcr.io/nvidia/vllm:26.04-py3
+vllm_image:   nvcr.io/nvidia/vllm:26.07-py3
diff --git a/agent-samples/model-servers/yaml/dual_48G_ada/vlm_server.yaml b/agent-samples/model-servers/yaml/dual_48G_ada/vlm_server.yaml
@@ -1,18 +1,18 @@
 # SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 # SPDX-License-Identifier: Apache-2.0
 
-# VLM server configuration — nvidia/Cosmos-Reason1-7B.
+# VLM server configuration — Cosmos3 Nano Reasoner.
 # Auto-discovered by xr-ai-launcher and passed as --config to vlm_server.
 # OpenAI-compatible API at http://localhost:8100/v1 (accepts image_url in messages)
 
-# HuggingFace model ID — any Qwen2.5-VL-compatible model is supported.
-# nvidia/Cosmos-Reason1-7B   — NVIDIA VLM; Apache 2.0 + NVIDIA Open Model License;
-#                              no HF login required; ~16 GB VRAM at BF16
-# nvidia/Cosmos-Reason2-8B   — stronger reasoning; gated — requires HF login + license
-model: nvidia/Cosmos-Reason1-7B
+# HuggingFace model ID. Standard vLLM serving of nvidia/Cosmos3-Nano loads the
+# Reasoner only; the Generator requires a separate `--omni` serving path.
+# Cosmos-Reason1 compatibility uses nvidia/Cosmos-Reason1-7B with the
+# cosmos_vlm client preset.
+model: nvidia/Cosmos3-Nano
 
-# HuggingFace token — required for gated models (e.g. Cosmos-Reason2-8B).
-# Accept license at https://huggingface.co/nvidia/Cosmos-Reason2-8B first.
+# HuggingFace token — optional for this public model and required for gated
+# alternatives.
 # !! Do not commit this file with a real token !!
 hf_token: ""
 
@@ -26,12 +26,12 @@ model_cache: ../../../../models
 # Single-image queries have short context so the KV budget is adequate.
 gpu_memory_utilization: 0.43
 enforce_eager: true
+async_scheduling: true
 max_videos_per_prompt: 0
-# 4 fits within the default max_model_len=8192 (each Qwen2.5-VL image is
-# ~1.5k tokens). Raise both together if multi-image turns get larger.
+# Raise max_model_len too if multi-image turns outgrow the context window.
 max_images_per_prompt: 4
 
 # vLLM runtime: pip (this wrapper's venv) or docker (NGC nvcr.io/nvidia/vllm).
 # See services/vlm-server/vlm_server.yaml for the full toggle docs.
 vllm_backend: docker
-vllm_image:   nvcr.io/nvidia/vllm:26.04-py3
+vllm_image:   nvcr.io/nvidia/vllm:26.07-py3
diff --git a/agent-samples/model-servers/yaml/spark/vlm_server.yaml b/agent-samples/model-servers/yaml/spark/vlm_server.yaml
@@ -5,22 +5,22 @@
 #
 # vlm is SECOND in load order: agent-llm → vlm.
 # free_at_startup ≈ 120 − 42 (agent-llm) − 2 (STT) ≈ 76 GiB.
-# 0.20 × 120 = 24 GiB cap: ~16 GiB Cosmos-7B weights + ~8 GiB KV.
+# 0.20 × 120 = 24 GiB cap: ~16 GiB Cosmos3 Reasoner weights + ~8 GiB KV.
 
-model: nvidia/Cosmos-Reason1-7B
+model: nvidia/Cosmos3-Nano
 hf_token: ""
 port: 8100
 host: "0.0.0.0"
 model_cache: ../../../../models
 gpu_memory_utilization: 0.20
 max_num_seqs: 4
 enforce_eager: true
+async_scheduling: true
 max_videos_per_prompt: 0
-# 4 fits within the default max_model_len=8192 (each Qwen2.5-VL image is
-# ~1.5k tokens). Raise both together if multi-image turns get larger.
+# Raise max_model_len too if multi-image turns outgrow the context window.
 max_images_per_prompt: 4
 
 # vLLM runtime: pip (this wrapper's venv) or docker (NGC nvcr.io/nvidia/vllm).
 # See services/vlm-server/vlm_server.yaml for the full toggle docs.
 vllm_backend: docker
-vllm_image:   nvcr.io/nvidia/vllm:26.04-py3
+vllm_image:   nvcr.io/nvidia/vllm:26.07-py3
diff --git a/agent-samples/simple-vlm-example/yaml/models.hosted.json b/agent-samples/simple-vlm-example/yaml/models.hosted.json
@@ -9,7 +9,7 @@
       "category": "vlm",
       "adapter": {
         "kind": "openai_compat",
-        "model_name": "nvidia/nemotron-nano-12b-v2-vl",
+        "model_name": "nvidia/cosmos3-nano-reasoner",
         "capabilities": {"streaming": true, "vision": true}
       },
       "endpoint": {

diff --git a/agent-samples/simple-vlm-example/yaml/models.local.json b/agent-samples/simple-vlm-example/yaml/models.local.json
@@ -6,7 +6,7 @@
       "deployment": {"ownership": "managed", "service": "stt"}
     },
     "vlm": {
-      "adapter": {"preset": "cosmos_vlm"},
+      "adapter": {"preset": "cosmos3_nano_reasoner"},
       "endpoint": {"base_url": "http://localhost:8100", "readiness": "health"},
       "deployment": {"ownership": "managed", "service": "vlm"}
     },

diff --git a/agent-samples/simple-vlm-example/yaml/vlm_server.yaml b/agent-samples/simple-vlm-example/yaml/vlm_server.yaml
@@ -4,11 +4,12 @@
 # vlm-server configuration for vlm-agent.
 # Auto-discovered by the launcher and passed as --config to vlm_server.
 
-model:       nvidia/Cosmos-Reason1-7B
+model:       nvidia/Cosmos3-Nano
 port:        8100
 model_cache: ../../../models
+async_scheduling: true
 
 # vLLM runtime: pip (this wrapper's venv) or docker (NGC nvcr.io/nvidia/vllm).
 # See services/vlm-server/vlm_server.yaml for the full toggle docs.
 vllm_backend: docker
-vllm_image:   nvcr.io/nvidia/vllm:26.04-py3
+vllm_image:   nvcr.io/nvidia/vllm:26.07-py3
diff --git a/agent-samples/xr-render-demo/yaml/models.nim.yaml b/agent-samples/xr-render-demo/yaml/models.nim.yaml
@@ -38,16 +38,14 @@ agent_llm:
     tool_calls: true
     reasoning:  true
 
-# Visual question answering used by native vision functions.
-# nvidia/cosmos-reason1-7b mirrors the local default (services/vlm-server
-# also runs Cosmos-Reason1-7B), so the render-demo's ask_image prompts behave
-# the same on NIM as locally. Swap to any hosted vision model on
-# build.nvidia.com if you prefer.
+# Visual question answering used by native vision functions. The hosted model
+# is the Cosmos3 Nano Reasoner, matching the role of the local Reasoner-only
+# vLLM service; it is not the separate Cosmos3 Generator.
 vlm:
   kind:        openai_compat
   category:    vlm
   base_url:    https://integrate.api.nvidia.com
-  model_name:  nvidia/cosmos-reason1-7b
+  model_name:  nvidia/cosmos3-nano-reasoner
   api_key_env: NGC_API_KEY
   health_check: false
   capabilities:

diff --git a/agent-samples/xr-render-demo/yaml/models.yaml b/agent-samples/xr-render-demo/yaml/models.yaml
@@ -23,5 +23,5 @@ tts:
   base_url: http://localhost:8105
 
 vlm:
-  kind: preset:cosmos_vlm
+  kind: preset:cosmos3_nano_reasoner
   base_url: http://localhost:8100
diff --git a/agent-sdk/xr-ai-models/README.md b/agent-sdk/xr-ai-models/README.md
@@ -57,7 +57,8 @@ Built-in presets — see `xr_ai_models/presets/`:
 
 | Preset | Service it targets | Notes |
 |---|---|---|
-| `cosmos_vlm`     | vlm-server               | image + video; `enable_thinking=false` by default. Video requires vlm-server's `max_videos_per_prompt >= 1` |
+| `cosmos3_nano_reasoner` | vlm-server          | default; Cosmos3 Nano text-output Reasoner, image + video; video requires `max_videos_per_prompt >= 1` |
+| `cosmos_vlm`     | vlm-server               | Cosmos-Reason1 compatibility option; image + video; `enable_thinking=false` by default; video requires `max_videos_per_prompt >= 1` |
 | `llama_nemotron` | llama-nemotron-llm-server | OpenAI tool calling via llama3_json (server-side) |
 | `nemotron3_nano` | nemotron3-nano-llm-server | reasoning field: `reasoning` |
 | `nemotron_omni`  | nemotron-omni-llm-server  | reasoning field: `reasoning_content`, vision + video |
@@ -186,7 +187,7 @@ are a profile change:
       "category": "vlm",
       "adapter": {
         "kind": "openai_compat",
-        "model_name": "nvidia/cosmos-reason1-7b"
+        "model_name": "nvidia/cosmos3-nano-reasoner"
       },
       "endpoint": {
         "base_url": "https://integrate.api.nvidia.com",

diff --git a/agent-sdk/xr-ai-models/xr_ai_models/presets/__init__.py b/agent-sdk/xr-ai-models/xr_ai_models/presets/__init__.py
@@ -13,6 +13,7 @@
 from copy import deepcopy
 from typing import Any
 
+from .cosmos3_nano_reasoner import COSMOS3_NANO_REASONER
 from .cosmos_vlm      import COSMOS_VLM
 from .llama_nemotron  import LLAMA_NEMOTRON
 from .magpie_tts      import MAGPIE_TTS
@@ -24,6 +25,7 @@
 
 
 _PRESETS: dict[str, dict[str, Any]] = {
+    "cosmos3_nano_reasoner": COSMOS3_NANO_REASONER,
     "cosmos_vlm":     COSMOS_VLM,
     "llama_nemotron": LLAMA_NEMOTRON,
     "magpie_tts":     MAGPIE_TTS,