fix(inference): update DeepSeek V4 Flash recipe and default DGX Station to DeepSeek V4 Flash

ci/platform-matrix.json

@@ -64,7 +64,7 @@
       "status": "deferred",
       "prd_priority": "P1",
       "ci_tested": false,
-      "notes": "The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Onboard path not yet validated end-to-end on the hardware; vLLM has a placeholder default model defined for this host class (`Qwen/Qwen3.6-27B-FP8`) that will move out of `deferred` once the hardware run is signed off."
+      "notes": "The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Onboard path not yet validated end-to-end on the hardware; vLLM defaults to `deepseek-ai/DeepSeek-V4-Flash` for this host class and will move out of `deferred` once the hardware run is signed off."
     },
     {
       "name": "NVIDIA RTX (consumer and Pro workstation GPUs)",
@@ -141,7 +141,7 @@
       "name": "Local vLLM (managed install/start)",
       "status": "caveated",
       "endpoint_type": "Local OpenAI-compatible",
-      "notes": "Appears by default on DGX Spark and DGX Station. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pulls or starts the stable NGC vLLM container for each host profile. See `src/lib/inference/vllm.ts:55,177` for the pins. DGX Spark and DGX Station use `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use `nvcr.io/nvidia/vllm:26.03.post1-py3`. Validated defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `Qwen/Qwen3.6-27B-FP8`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing."
+      "notes": "Appears by default on DGX Spark and DGX Station. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pulls or starts the stable NGC vLLM container for each host profile. See `src/lib/inference/vllm.ts:55,177` for the pins. DGX Spark and DGX Station use `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use `nvcr.io/nvidia/vllm:26.03.post1-py3`. Validated defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing."
     }
   ],
 

docs/get-started/quickstart.mdx

@@ -78,7 +78,7 @@ curl -fsSL https://www.nvidia.com/nemoclaw.sh | bash
 
 On DGX Spark, DGX Station, and Windows WSL, an interactive installer offers express install after you accept the third-party software notice.
 Express install switches onboarding to non-interactive mode, allows `sudo` password prompts for required host changes, and selects the managed local inference path for that platform.
-On DGX Spark and DGX Station, express install selects managed vLLM; DGX Spark uses `qwen3.6-35b-a3b-nvfp4` by default.
+On DGX Spark and DGX Station, express install selects managed vLLM; DGX Spark uses `qwen3.6-35b-a3b-nvfp4` by default, and DGX Station uses `deepseek-v4-flash` by default.
 Unless `NEMOCLAW_POLICY_TIER` is set, it applies sandbox policy in `suggested` mode with the `balanced` tier by default.
 That tier uses the base sandbox policy plus supported package, model, web-search, local-inference, and read-only weather presets.
 On DGX Spark, express install uses `my-spark-assistant` as the sandbox name unless `NEMOCLAW_SANDBOX_NAME` is already set.

docs/inference/inference-options.mdx

@@ -51,7 +51,7 @@ NemoClaw uses provider-specific local tokens for those routes, and rebuilds of l
 | Local Ollama | Tested with limitations | Local Ollama API | Available when Ollama is installed or running on the host. Validated default models: `qwen3.6:35b` (high VRAM), `nemotron-3-nano:30b` (medium VRAM), `qwen3.5:9b` (low VRAM fallback). |
 | Local NVIDIA NIM | Experimental | Local OpenAI-compatible | Requires `NEMOCLAW_EXPERIMENTAL=1` and a NIM-capable NVIDIA GPU. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence with `assertCdiNvidiaGpuSpecPresent`, `src/lib/onboard.ts:1586`). NIM images pull from `nvcr.io` and require NGC registry login. NemoClaw gates this path behind the experimental flag because it does not auto-select a NIM image for the host today. You must explicitly pick from the validated image list. Managed vLLM has host-specific default models and is not gated on the same boxes. Validated images referenced in `src/lib/inference/config.ts` and `nemoclaw/src/index.ts`: `nvidia/nemotron-3-super-120b-a12b` (default cloud model), `nvidia/nemotron-3-nano-30b-a3b`, `nvidia/llama-3.3-nemotron-super-49b-v1.5`. |
 | Local vLLM (already running) | Tested with limitations | Local OpenAI-compatible | Appears in the onboarding menu when NemoClaw detects a server already on `localhost:8000`. No flag required. Model is whatever the existing server serves. |
-| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and DGX Station. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pulls or starts the stable NGC vLLM container for each host profile. See `src/lib/inference/vllm.ts:55,177` for the pins. DGX Spark and DGX Station use `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use `nvcr.io/nvidia/vllm:26.03.post1-py3`. Validated defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `Qwen/Qwen3.6-27B-FP8`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. |
+| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and DGX Station. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pulls or starts the stable NGC vLLM container for each host profile. See `src/lib/inference/vllm.ts:55,177` for the pins. DGX Spark and DGX Station use `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use `nvcr.io/nvidia/vllm:26.03.post1-py3`. Validated defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. |
 {/* provider-status:end */}
 
 ## Provider Options
@@ -318,7 +318,7 @@ Managed vLLM uses these profiles:
 | Host profile | Default model |
 |---|---|
 | DGX Spark | `nvidia/Qwen3.6-35B-A3B-NVFP4` |
-| DGX Station | `Qwen/Qwen3.6-27B-FP8` |
+| DGX Station | `deepseek-ai/DeepSeek-V4-Flash` |
 | Linux with an NVIDIA GPU | `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8` |
 
 <Note>
@@ -356,10 +356,10 @@ Recognized slugs are:
 
 | Slug | Hugging Face model | Notes |
 |---|---|---|
-| `qwen3.6-27b` | `Qwen/Qwen3.6-27B-FP8` | Default on the DGX Station profile |
+| `qwen3.6-27b` | `Qwen/Qwen3.6-27B-FP8` | Supported override |
 | `qwen3.6-35b-a3b-nvfp4` | `nvidia/Qwen3.6-35B-A3B-NVFP4` | Default on the DGX Spark profile |
 | `nemotron-3-nano-4b` | `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8` | Default on the generic Linux + NVIDIA GPU profile |
-| `deepseek-v4-flash` | `deepseek-ai/DeepSeek-V4-Flash` | Supported override |
+| `deepseek-v4-flash` | `deepseek-ai/DeepSeek-V4-Flash` | Default on the DGX Station profile |
 | `deepseek-r1-distill-70b` | `deepseek-ai/DeepSeek-R1-Distill-Llama-70B` | Gated. Requires Hugging Face license acceptance |
 
 The slug is case-insensitive; the full Hugging Face id is also accepted.

docs/reference/platform-support.mdx

@@ -82,7 +82,7 @@ For the onboarding-time supported set without deferred rows, refer to [Prerequis
 | macOS (Apple Silicon) | Colima, Docker Desktop | Tested with limitations | P0 | Yes | Start the container runtime (Colima or Docker Desktop) before running the installer. Xcode Command Line Tools (`xcode-select --install`) are typically required for Node native modules during install. NemoClaw recommends them but does not enforce them during preflight. |
 | DGX Spark | Docker | Tested | P1 | Yes | Use the standard installer and `$$nemoclaw onboard`. For an end-to-end walkthrough with local inference, see the [NVIDIA Spark playbook](https://build.nvidia.com/spark/nemoclaw). |
 | Windows WSL2 | Docker Desktop (WSL backend) | Tested with limitations | P1 | No | Requires WSL2 with Docker Desktop backend. |
-| DGX Station | Docker | Deferred | P1 | No | The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Onboard path not yet validated end-to-end on the hardware; vLLM has a placeholder default model defined for this host class (`Qwen/Qwen3.6-27B-FP8`) that will move out of `deferred` once the hardware run is signed off. |
+| DGX Station | Docker | Deferred | P1 | No | The PRD marks this platform as P1. Workstation form-factor with NVIDIA GPUs and the same Docker + NVIDIA Container Toolkit + CDI requirements as DGX Spark. Onboard path not yet validated end-to-end on the hardware; vLLM defaults to `deepseek-ai/DeepSeek-V4-Flash` for this host class and will move out of `deferred` once the hardware run is signed off. |
 | NVIDIA RTX (consumer and Pro workstation GPUs) | Docker | Deferred | P1 | No | The PRD marks this platform as P1. Covers RTX consumer cards and RTX Pro workstation cards on Linux hosts that meet the generic-Linux-GPU requirements (NVIDIA Container Toolkit + CDI present). The provider menu emits managed vLLM behind `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm` for this host class today; the end-to-end onboard path on this hardware is not yet validated in CI. |
 {/* platform-matrix-full:end */}
 
@@ -103,7 +103,7 @@ NemoClaw routes inference through the OpenShell gateway. Each row below is a pro
 | Local Ollama | Tested with limitations | Local Ollama API | Available when Ollama is installed or running on the host. Validated default models: `qwen3.6:35b` (high VRAM), `nemotron-3-nano:30b` (medium VRAM), `qwen3.5:9b` (low VRAM fallback). |
 | Local NVIDIA NIM | Experimental | Local OpenAI-compatible | Requires `NEMOCLAW_EXPERIMENTAL=1` and a NIM-capable NVIDIA GPU. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence with `assertCdiNvidiaGpuSpecPresent`, `src/lib/onboard.ts:1586`). NIM images pull from `nvcr.io` and require NGC registry login. NemoClaw gates this path behind the experimental flag because it does not auto-select a NIM image for the host today. You must explicitly pick from the validated image list. Managed vLLM has host-specific default models and is not gated on the same boxes. Validated images referenced in `src/lib/inference/config.ts` and `nemoclaw/src/index.ts`: `nvidia/nemotron-3-super-120b-a12b` (default cloud model), `nvidia/nemotron-3-nano-30b-a3b`, `nvidia/llama-3.3-nemotron-super-49b-v1.5`. |
 | Local vLLM (already running) | Tested with limitations | Local OpenAI-compatible | Appears in the onboarding menu when NemoClaw detects a server already on `localhost:8000`. No flag required. Model is whatever the existing server serves. |
-| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and DGX Station. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pulls or starts the stable NGC vLLM container for each host profile. See `src/lib/inference/vllm.ts:55,177` for the pins. DGX Spark and DGX Station use `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use `nvcr.io/nvidia/vllm:26.03.post1-py3`. Validated defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `Qwen/Qwen3.6-27B-FP8`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. |
+| Local vLLM (managed install/start) | Tested with limitations | Local OpenAI-compatible | Appears by default on DGX Spark and DGX Station. Generic Linux NVIDIA GPU hosts require `NEMOCLAW_EXPERIMENTAL=1` or `NEMOCLAW_PROVIDER=install-vllm`. Host must have the NVIDIA Container Toolkit installed and a CDI spec present (`onboard` asserts CDI presence). NemoClaw pulls or starts the stable NGC vLLM container for each host profile. See `src/lib/inference/vllm.ts:55,177` for the pins. DGX Spark and DGX Station use `nvcr.io/nvidia/vllm:26.05.post1-py3`; generic Linux NVIDIA GPU hosts use `nvcr.io/nvidia/vllm:26.03.post1-py3`. Validated defaults are listed in `src/lib/inference/vllm-models.ts`: DGX Spark uses `nvidia/Qwen3.6-35B-A3B-NVFP4`, DGX Station uses `deepseek-ai/DeepSeek-V4-Flash`, and Linux NVIDIA GPU uses `nvidia/NVIDIA-Nemotron-3-Nano-4B-FP8`. Image pulls require NGC registry login (`docker login nvcr.io`); onboard prompts for the NGC API key when authentication is missing. |
 {/* provider-status-full:end */}
 
 ## Messaging Integrations

src/lib/inference/model-prompts.test.ts

@@ -235,7 +235,7 @@ describe("promptVllmModel", () => {
   const sparkDefault = VLLM_MODELS.find((m) => m.envValue === "qwen3.6-35b-a3b-nvfp4")!;
   const gatedModel = VLLM_MODELS.find((m) => m.envValue === "deepseek-r1-distill-70b")!;
   const stationModels = modelsForPlatform("station");
-  const stationDefault = VLLM_MODELS.find((m) => m.envValue === "qwen3.6-27b")!;
+  const stationDefault = VLLM_MODELS.find((m) => m.envValue === "deepseek-v4-flash")!;
 
   it("returns the profile default when the user presses Enter", async () => {
     const promptFn = promptSequence([""]);

src/lib/inference/vllm-models.test.ts

@@ -170,9 +170,7 @@ describe("vllm model registry", () => {
     expect(cmd).toContain("--no-disable-hybrid-kv-cache-manager");
     expect(cmd).toContain("--disable-uvicorn-access-log");
     expect(cmd).toContain("--max-cudagraph-capture-size 128");
-    expect(cmd).toContain(
-      `--speculative-config '{"method":"mtp","num_speculative_tokens":3,"rejection_sample_method":"synthetic","synthetic_acceptance_length":3}'`,
-    );
+    expect(cmd).toContain(`--speculative-config '{"method":"mtp","num_speculative_tokens":3}'`);
     expect(cmd).toContain("--max-model-len 1048576");
     expect(cmd).toContain("--max-num-batched-tokens 8192");
     expect(cmd).toContain("--max-num-seqs 16");

src/lib/inference/vllm-models.ts

@@ -141,7 +141,7 @@ export const VLLM_MODELS: readonly VllmModelDef[] = [
       "--max-cudagraph-capture-size",
       "128",
       "--speculative-config",
-      `'{"method":"mtp","num_speculative_tokens":3,"rejection_sample_method":"synthetic","synthetic_acceptance_length":3}'`,
+      `'{"method":"mtp","num_speculative_tokens":3}'`,
       "--max-num-batched-tokens",
       "8192",
       "--max-num-seqs",
@@ -214,8 +214,8 @@ export const VLLM_EXTRA_ARGS_ENV = "NEMOCLAW_VLLM_EXTRA_ARGS_JSON";
 /**
  * Look up the requested express-vLLM model from `NEMOCLAW_VLLM_MODEL`.
  * Returns `null` when the env var is empty so the caller can fall back to
- * the per-platform profile default (Station prefers Qwen3.6-27B, Spark the
- * Qwen3.6-35B-A3B NVFP4 checkpoint, and the generic Linux profile prefers
+ * the per-platform profile default (Station prefers DeepSeek V4 Flash, Spark
+ * the Qwen3.6-35B-A3B NVFP4 checkpoint, and the generic Linux profile prefers
  * Nemotron-Nano-4B for VRAM headroom).
  *
  * Match is case-insensitive against either the `envValue` slug or the full

src/lib/inference/vllm.test.ts

@@ -34,14 +34,14 @@ describe("vLLM profile detection", () => {
     vi.clearAllMocks();
   });
 
-  it("uses Qwen3.6 27B and the 26.05.post1 NGC image on DGX Station", () => {
+  it("uses DeepSeek V4 Flash and the 26.05.post1 NGC image on DGX Station", () => {
     const profile = detectVllmProfile({ platform: "station", type: "nvidia" });
 
     expect(profile).not.toBeNull();
     expect(profile!.name).toBe("DGX Station");
     expect(profile!.image).toBe("nvcr.io/nvidia/vllm:26.05.post1-py3");
-    expect(profile!.defaultModel.id).toBe("Qwen/Qwen3.6-27B-FP8");
-    expect(profile!.defaultModel.envValue).toBe("qwen3.6-27b");
+    expect(profile!.defaultModel.id).toBe("deepseek-ai/DeepSeek-V4-Flash");
+    expect(profile!.defaultModel.envValue).toBe("deepseek-v4-flash");
   });
 
   it("keeps DGX Spark on the Qwen3.6 35B NVFP4 default", () => {

src/lib/inference/vllm.ts

@@ -62,9 +62,9 @@ function nemotronNanoModel(): VllmModelDef {
   return match;
 }
 
-function qwen27bFP8Model(): VllmModelDef {
-  const match = VLLM_MODELS.find((m) => m.envValue === "qwen3.6-27b");
-  if (!match) throw new Error("vllm-models registry is missing the qwen3.6-27b entry");
+function deepseekV4FlashModel(): VllmModelDef {
+  const match = VLLM_MODELS.find((m) => m.envValue === "deepseek-v4-flash");
+  if (!match) throw new Error("vllm-models registry is missing the deepseek-v4-flash entry");
   return match;
 }
 
@@ -144,7 +144,7 @@ const STATION_PROFILE: VllmProfile = {
   name: "DGX Station",
   platform: "station",
   image: VLLM_IMAGES.ngc2605Post1,
-  defaultModel: qwen27bFP8Model(),
+  defaultModel: deepseekV4FlashModel(),
   containerName: "nemoclaw-vllm",
   dockerRunFlags: SPARK_PROFILE.dockerRunFlags,
   buildDockerRunFlags: () => {

test/detect-vllm-profile.test.ts

@@ -29,8 +29,8 @@ describe("detectVllmProfile", () => {
     expect(profile).not.toBeNull();
     expect(profile!.name).toBe("DGX Station");
     expect(profile!.image).toBe("nvcr.io/nvidia/vllm:26.05.post1-py3");
-    expect(profile!.defaultModel.id).toBe("Qwen/Qwen3.6-27B-FP8");
-    expect(profile!.defaultModel.envValue).toBe("qwen3.6-27b");
+    expect(profile!.defaultModel.id).toBe("deepseek-ai/DeepSeek-V4-Flash");
+    expect(profile!.defaultModel.envValue).toBe("deepseek-v4-flash");
   });
 
   it("returns the generic Linux profile for non-Spark/Station NVIDIA hosts", () => {