DGX OpenClaw Stack

A self-hosted OpenClaw agent on one NVIDIA GB10 box (DGX Spark / ASUS Ascent GX10), brought up with docker compose.

The model: DeepSeek-V4.1-Flash, 284B parameters, on one desktop box. It runs on Strata's engine in a ~1.75-bit quantisation: the experts the model routes to most sit on the GPU, the rest is computed on the CPU straight from the files. On the reference GX10 it serves the Discord bot, the MeshCore mesh, and Claude Code on the LAN, offline. It comes back by itself after a reboot, stays below 80 °C under sustained load, and its limits allow agent runs of up to a week. It is not fast (code 20–33 tok/s, prose 15–17, one request at a time), but it is the largest model this box runs, and it runs reliably (openclaw-strata/README.md).

Around it: bge-m3 embeddings for memory, SearxNG web search, faster-whisper for voice notes, and OpenClaw's Web UI, agents and Discord bot. The quick start needs no engine build: Nemotron 3.5 Super on vLLM, one docker compose up; Gemma 4 and Mistral Small 4 are other vLLM options. The value is the wiring: config-init writes openclaw.json and the agent workspaces from .env on every docker compose up, so the setup stays the same across onboarding re-runs, restarts and OpenClaw upgrades.

What runs

Service	What it does	Host port	Starts with
strata	Chat model: DeepSeek-V4.1-Flash (ISTA GSQ-RCO IQ1_M) on Strata's strata-dsv41 engine, built on the host; tools, thinking, DSpark drafting, text only	127.0.0.1:8097	compose/strata.yml, started by llm-ensure.sh after the stack
strata-router	Claude Code's endpoint for it: keeps the auto-mode classifier inside its 60 s limit	127.0.0.1:8096	compose/strata-router.yml
vllm-llm	Quick-start chat model: Nemotron 3.5 Super 120B-A12B VL NVFP4 on vllm/vllm-openai:v0.32.0, with the LiLiCorr dflash drafter (vision, tools, medium-effort thinking); parked by compose/strata-llm.yml	127.0.0.1:8004	default
vllm-embedding	BAAI/bge-m3 embeddings for memory search	none (8005 inside)	default
searxng	Web search backend for the web_search tool	none (8080 inside)	default
openclaw-stt-whisper	Speech-to-text, faster-whisper-large-v3-turbo, Hungarian and English detected per request	127.0.0.1:8093	default
openclaw-config-init	Writes openclaw.json and the workspaces from .env, then exits	none	default
openclaw-gateway	OpenClaw: Web UI, agents, Discord	18789	default
openclaw-cli	Runs openclaw ... commands; shares the gateway's network namespace	none	default
openclaw-browser	Chromium with named login profiles, driven over CDP; noVNC for logging in	127.0.0.1:9220, 9222-9241, 5901	profile browser
openclaw-python-sandbox	Python over MCP, one kernel per session, plus yt-dlp, ffmpeg and a static preview server	127.0.0.1:8094, 8095	profile python
openclaw-meshcore-bridge	LoRa mesh access through a USB MeshCore companion radio	none	profile meshcore

.env.example turns on COMPOSE_PROFILES=browser,python. Services reach each other by service name on the compose network.

The gateway image is openclaw-base-ext: the upstream OpenClaw image plus ffmpeg, git, curl and four local plugins (people-context, claim-guard, dm-guild-gate, mcp-image-saver). The upstream entrypoint is kept, and the upstream code is not patched.

Optional extras

Extra compose files go into COMPOSE_FILE in .env, for example COMPOSE_FILE=docker-compose.yml:compose/gemma4.yml. Each file's header says what else it needs.

File	What it adds
compose/gemma4.yml	Gemma 4 instead of Nemotron, on the same service and port (26B-A4B MoE or 31B dense)
compose/mistral.yml	Mistral Small 4 119B-A6B NVFP4 instead of Nemotron
compose/tts-fish.yml	Text-to-speech with Fish Audio S2 Pro (80+ languages, voice cloning). Off by default
compose/claw-img.yml	A separate Discord bot with /claw-img, /claw-video and /claw-music slash commands that call the ComfyUI bridge directly. Off by default
compose/strata.yml	DeepSeek-V4.1-Flash (284B MoE, ~1.75 bit) on Strata's engine: hot experts on the GPU, the rest on the CPU. One request at a time
compose/strata-llm.yml	That DeepSeek as OpenClaw's chat model instead of Nemotron
compose/strata-router.yml	Claude Code against that DeepSeek, auto mode included (below)

A model file needs LLM_MODEL_ID set to the id it serves, because config-init registers that id in OpenClaw.

Claude Code on the Spark, offline. With compose/strata.yml and compose/strata-router.yml, Claude Code on any machine in the LAN runs against the DeepSeek on the GB10: no request leaves the network. Code comes at 20–33 tok/s, prose at 15–17, and a resumed turn reads only its new tokens. The router keeps Claude Code's auto-mode safety classifier inside its fixed 60 s limit, so long unattended runs work. Setup and the sclaude launchers: openclaw-strata-router/README.md; the server: openclaw-strata/README.md.

The ComfyUI image, video and music bridge has its own compose file in openclaw-image-comfyui/. It talks to a ComfyUI install you run yourself, and the repo ships no model weights. The agent's media tools stay off (OPENCLAW_DISCORD_MEDIA_TOOLS_OFF) until you turn them on.

Hardware and measured numbers

The defaults are sized for a GB10: one 128 GB memory pool shared by the CPU and the GPU, so each vLLM server gets its share through --gpu-memory-utilization. Two things matter on this box:

• Install the GPU clock cap. Under sustained load an uncapped GB10 can power itself off with nothing in the journal, at about 94 °C, well below the ACPI trip point. With the clock locked at 2200 MHz, a 25-minute prefill stress test stayed at 87 °C or below, and decode speed did not drop. Details: docs/reference/gb10-power-thermal.md.
• Do not overfill the pool. When it runs dry, the host stalls instead of killing a process. Nemotron takes about 70 GiB at the default share, so the Fish TTS and a ComfyUI install do not fit next to it.

Both, and everything else the stack needs outside Docker (the thermal and memory guards, the job that brings the chat model back after a reboot, kernel settings that keep the box reachable), come from one script: sudo bash scripts/gb10-host-setup.sh system, ... resilience, then bash scripts/gb10-host-setup.sh cron. check compares the host with the repo and changes nothing (docs/reference/host-setup.md).

Measured on the reference GX10. The setup for each number is in docs/reference/llm-stack.md and the compose file headers.

	DeepSeek-V4.1-Flash on Strata	Nemotron 3.5 Super (quick start)	Mistral Small 4	Gemma 4 26B-A4B MoE	Gemma 4 31B dense
Parameters	284B MoE	120B-A12B	119B-A6B	26B-A4B	31B
Weights	121 GB IQ1_M + 111 GB Engram table, on disk	~80 GB download	~66 GiB	~16.5 GB	~17 GB
Memory	~80–88 GB of experts on the GPU (~11k of 15.7k), the rest read from the files by the CPU	0.70 share (~70 GiB)	0.65 (~78 GiB process)	0.50	0.50
Decode, 1 stream	code 20–33 tok/s (DSpark), prose 15–17 tok/s	26.2 tok/s English, 16.6 tok/s Hungarian (the drafter predicts Hungarian poorly); 19.8 tok/s with the clock cap	~30 tok/s	~25–52 tok/s (depends on the vLLM build)	~6.9 tok/s
Decode, 4 streams	one request at a time; the rest queue	8.8 tok/s each	19 tok/s each	n/a	n/a
Prefill	~450 tok/s; a resumed conversation reads only its new tokens	~2560 tok/s	~1900 tok/s	n/a	n/a
Engine	Strata 0.2.7 (strata-dsv41), built on the host	vLLM v0.32.0	vLLM v0.31.0	vLLM v0.32.0	vLLM v0.32.0

DeepSeek is the reference host's model since 2026-10-28: the bot, the mesh and Claude Code all run on it. Its numbers are measured beside the rest of the stack, with the engine pinned to the A725 cores (acpitz below 80 °C). Strata measures this quantisation at +9–10% perplexity on code and agent work against the full model. The Gemma numbers were measured on older vLLM releases with a patched image and have not been measured again on stock v0.32. Of the vLLM models, Nemotron gave the most reliable facts and tool use in the repo's Discord agent evals, which is why it is the quick-start default. Its Hungarian is weaker than Mistral's or Gemma's; config-init sets temperature 0.6 for it, which keeps other scripts out of Hungarian replies.

On other hardware, everything except vllm-llm carries over. To use a different model, a remote vLLM or a cloud endpoint, see docs/CUSTOMIZATION.md and LLM_BASE_URL / EMBED_BASE_URL in docs/reference/configuration.md.

Quickstart

git clone https://github.com/chestercs/dgx-openclaw-stack.git
cd dgx-openclaw-stack

./bootstrap.sh            # creates .env, fills the secrets, asks for the HF token
docker compose up -d      # first start downloads ~80 GB of weights

The first start takes two phases, on purpose. Before onboarding there is no openclaw.json, so config-init has nothing to change and the gateway restarts with Missing config. OpenClaw wants you to onboard explicitly, so that you choose the gateway token. Onboard as described in SETUP.md, then let config-init wire the new config:

docker compose up -d --force-recreate openclaw-config-init openclaw-gateway openclaw-cli
curl -sS http://127.0.0.1:18789/healthz

Always recreate those three together. openclaw-cli lives in the gateway's network namespace and loses its network when the gateway is recreated without it.

The full walkthrough, with sanity checks and day-to-day operation, is in SETUP.md.

Then move to DeepSeek. The quick start runs on Nemotron. For DeepSeek, build Strata's engine on the host, download the model (~232 GB) and build its pack, then add compose/strata.yml and compose/strata-llm.yml to COMPOSE_FILE: openclaw-strata/README.md and SETUP.md.

Licences

The repo's code is MIT. Model weights keep their upstream licences. They are downloaded on first start and are not part of the repo:

• Nemotron 3.5 Super and its drafter: NVIDIA's licence on the model card.
• bge-m3: MIT. Whisper turbo (deepdml/faster-whisper-large-v3-turbo-ct2): MIT.
• Gemma 4 (compose/gemma4.yml): Apache 2.0. Mistral Small 4 (compose/mistral.yml): Apache 2.0.
• DeepSeek-V4.1-Flash and its quantisations (compose/strata.yml): the licences on the model cards; you download them yourself. Strata, which the container mounts from your own build: MIT.
• Fish Audio S2 Pro, used only by the optional compose/tts-fish.yml: the Fish Audio Research License, non-commercial use only. Building that image downloads the weights and counts as accepting the licence. The default stack does not include it. The wrapper code in openclaw-tts-fish/ is MIT. For the commercial route, see openclaw-tts-fish/README.md.

Documentation

• SETUP.md: first boot, onboarding, checks, daily operation.
• docs/reference/configuration.md: every .env knob, with its default and effect.
• docs/ARCHITECTURE.md: how the pieces fit together.
• docs/CUSTOMIZATION.md: other models, remote backends, other hardware.
• docs/TROUBLESHOOTING.md: known failure modes.
• docs/discord-bot-setup.md: creating the Discord bot.
• docs/OPENCLAW_UPGRADE.md: upgrading the OpenClaw image.
• docs/reference/host-setup.md: what runs on the host outside Docker, and the script that installs and checks it.
• openclaw-strata/README.md and openclaw-strata-router/README.md: DeepSeek-V4.1-Flash on Strata, and Claude Code against it, offline.
• docs/reference/README.md: deeper reference (LLM stack, GB10 thermals, Discord agent, browser, sandbox, MeshCore).
• CONTRIBUTING.md and CHANGELOG.md.

Issues that are not about this stack belong upstream: vllm-project/vllm or openclaw/openclaw.