# dgx-spark-qwen38

> dgx-spark-qwen38 is an open-source installer that turns an NVIDIA DGX Spark, or any GB10 machine (ASUS, Dell, HP, Lenovo, MSI, GIGABYTE, Acer), into a private AI server in one command: measured model lanes, OpenAI and Anthropic APIs, and a local cockpit. No prompt, image or video leaves the box.

Install, on the box: `curl -fsSL https://raw.githubusercontent.com/hasso5703/dgx-spark-qwen38/main/get.sh | bash`. The same command updates it. Open source, MIT license. Version v1.40.2 (2026-10-24).
The project's own numbers are measured on an ASUS Ascent GX10 with 128 GB, its reference box, with their method in the repo. Numbers from owners' reports are attributed and linked.

## Lanes (models), measured
- Qwen3.8-27B (Qwen): 71.4 tok/s, greedy median, one request. 135 to 148 tok/s in all at 8 streams. A window of a million tokens. GSM8K 94.0% on 200 problems. NVFP4 with a DFlash2 drafter, on SGLang. Source: https://github.com/hasso5703/dgx-spark-qwen38/blob/main/BENCHMARKS.md
- Qwen3.8-Flash-Next 176B (Qwen): 47.9 tok/s, on code, with FLASH_TIER=context. A 262K-token window. GSM8K 97.41% on 1,314 problems. HumanEval 95.73%, MMLU 90.8%. NVFP4 with NEXTN speculation, on SGLang. Source: https://github.com/hasso5703/dgx-spark-qwen38/blob/main/docs/flash-lane.md
- Flash FP4-Local (NVIDIA): 55.4 tok/s, greedy median, at its default effort. 106.6 tok/s in all at 8 streams, at the lean effort. GSM8K 97.64% on 1,314 problems. About half the disk of the flash lane. on vLLM. Source: https://github.com/hasso5703/dgx-spark-qwen38/blob/main/docs/flash-local-lane.md
- Nemotron 3.5 Bolt (NVIDIA): 140 tok/s, rewriting code, one request, reasoning off. 201.2 tok/s in all at 16 streams. Reads images and video. The lane a 64 GB Spark serves first. with its drafter, on vLLM. Source: https://github.com/hasso5703/dgx-spark-qwen38/blob/main/docs/nemotron-lane.md
- Nemotron 3.5 Super VL (NVIDIA): 69.5 tok/s, rewriting code, one request, reasoning off. 81.7 tok/s in all at 4 streams. A window of a million tokens. Reads images and video. with its drafter, on vLLM. Source: https://github.com/hasso5703/dgx-spark-qwen38/blob/main/docs/nemotron-super-lane.md
- Ornith 1.5 122B (Ornith AI): 42.4 tok/s, greedy median, one request. 90.0 tok/s in all at 8 streams on five real jobs. GSM8K 98.8% on 250 problems, on a second GX10. A window of a million tokens. NVFP4 with its MTP head, on vLLM. Source: https://github.com/hasso5703/dgx-spark-qwen38/blob/main/docs/ornith-lane.md
- Kolibri-1 (Aleph Alpha): 48.0 tok/s, one request. 121.9 tok/s at 4 streams. 165.8 at 8, 202.0 at 16. Aleph Alpha's FP8 weights. FP8, on vLLM. Source: https://github.com/hasso5703/dgx-spark-qwen38/blob/main/docs/kolibri-lane.md
- Qwen-Image 2.1 (Qwen): 7.5 s, with its Turbo, per 1024×1024 image. 34.3 s at 40 steps without it. The Turbo fetches 32.5 GB at its first switch. On the same box as the text lanes. on SGLang. Source: https://github.com/hasso5703/dgx-spark-qwen38/blob/main/docs/image-lane.md
- MiniMax-H3 (MiniMax): 126.2 to 127.9 s, with the Turbo adapter, for 4 s of 480P video. 622.8 to 632.5 s at 50 steps without it. Sound made with the picture. From a sentence. on SGLang. Source: https://github.com/hasso5703/dgx-spark-qwen38/blob/main/docs/video-lane.md

## Pages
- [Every page in full, in Markdown](https://hasso5703.github.io/llms-full.txt): the text of all the pages in one file, with the date each one last changed
- [Home](https://hasso5703.github.io/): what it is, the nine lanes, the clients, the machines
- [What to run on a 64 GB DGX Spark](https://hasso5703.github.io/dgx-spark-64gb/): Nemotron 3.5 Bolt at 110.8 tok/s and Qwen3.8-27B at 67.2 tok/s at a 64 GB budget, what needs 128 GB
- [Nine models, measured on a DGX Spark](https://hasso5703.github.io/models/)
- [Run Qwen3.8-27B on a DGX Spark](https://hasso5703.github.io/models/qwen3-8-27b/): 71.4 tok/s, greedy median, one request
- [Run Qwen3.8-Flash-Next 176B on a DGX Spark](https://hasso5703.github.io/models/qwen3-8-flash-next-176b/): 47.9 tok/s, on code, with FLASH_TIER=context
- [Run Flash FP4-Local on a DGX Spark](https://hasso5703.github.io/models/flash-fp4-local/): 55.4 tok/s, greedy median, at its default effort
- [Run Nemotron 3.5 Bolt on a DGX Spark](https://hasso5703.github.io/models/nemotron-3-5-bolt/): 140 tok/s, rewriting code, one request, reasoning off
- [Run Nemotron 3.5 Super VL on a DGX Spark](https://hasso5703.github.io/models/nemotron-3-5-super-vl/): 69.5 tok/s, rewriting code, one request, reasoning off
- [Run Ornith 1.5 122B on a DGX Spark](https://hasso5703.github.io/models/ornith-1-5-122b/): 42.4 tok/s, greedy median, one request
- [Run Kolibri-1 on a DGX Spark](https://hasso5703.github.io/models/kolibri-1/): 48.0 tok/s, one request
- [Run Qwen-Image 2.1 on a DGX Spark](https://hasso5703.github.io/models/qwen-image-2-1/): 7.5 s, with its Turbo, per 1024×1024 image
- [Run MiniMax-H3 on a DGX Spark](https://hasso5703.github.io/models/minimax-h3/): 126.2 to 127.9 s, with the Turbo adapter, for 4 s of 480P video
- [Guides for DGX Spark owners](https://hasso5703.github.io/guides/)
- [How to run a local LLM on a DGX Spark](https://hasso5703.github.io/guides/run-local-llm-dgx-spark/): Ollama, NVIDIA's playbooks, an engine you tune or a one-command installer: how each runs an LLM on a DGX Spark, what it asks of you, and the speeds measured.
- [An OpenAI-compatible API on a DGX Spark](https://hasso5703.github.io/guides/openai-compatible-api-dgx-spark/): Point Open WebUI, Cursor, Continue, Zed or the OpenAI SDK at your DGX Spark: one base URL on port 30001 and an API key. Any model name works.
- [Can a DGX Spark run a 1M-token context?](https://hasso5703.github.io/guides/dgx-spark-1m-token-context/): A plain install serves Qwen3.8-27B with a 1,010,000-token window on a DGX Spark. The KV pool behind it, the other 1M lanes, prefill times and how to switch.
- [Claude Code with a local model on a DGX Spark](https://hasso5703.github.io/guides/claude-code-local-model-dgx-spark/): Run Claude Code on the model your DGX Spark serves: claude-local sets the endpoint, the key and the limits of the lane, and keeps the session off Anthropic.
- [Image and video generation on a DGX Spark](https://hasso5703.github.io/guides/image-and-video-generation-dgx-spark/): Qwen-Image 2.1 makes a 1024×1024 image in 7.5 s with its Turbo on a DGX Spark, and MiniMax-H3 4 s of video with its sound in about two minutes.
- [opencode with a local model on a DGX Spark](https://hasso5703.github.io/guides/opencode-local-model-dgx-spark/): The installer sets up opencode for your DGX Spark's model: an oc launcher that lifts the hidden 32,000-token output cap, limits fitted to the pool.
- [Speech recognition and text to speech on a DGX Spark](https://hasso5703.github.io/guides/speech-recognition-and-text-to-speech-dgx-spark/): Qwen3-ASR hears 30 languages and Kokoro speaks seven, on a DGX Spark's CPU beside any model, behind OpenAI-compatible audio routes.
- [Update the AI stack of a DGX Spark without breaking it](https://hasso5703.github.io/guides/update-dgx-spark-ai-stack/): One command installs and updates the AI stack of a DGX Spark: pinned images and checkpoints, your choices kept, the serving model proved.
- [Which inference engine should you run on a DGX Spark?](https://hasso5703.github.io/guides/sglang-vs-vllm-vs-llama-cpp-on-dgx-spark/): SGLang, vLLM and llama.cpp on one DGX Spark: the drafter and the content matter more than the engine. Code: 28-32 tok/s SGLang with DSpark, 25-26 llama.cpp.
- [Why does my DGX Spark freeze when a model loads?](https://hasso5703.github.io/guides/dgx-spark-freezes-when-loading-a-model/): SGLang misses 25-40 GB of transient allocations on a DGX Spark, and a pool pushed to zero freezes it. Why it happens, how to see it coming and how to avoid it.
- [Why is Ollama slow on my DGX Spark?](https://hasso5703.github.io/guides/ollama-slow-on-dgx-spark/): Owners report Ollama slow on DGX Spark. What they found, what to check, and one same-harness test: 60.7 tok/s on vLLM against 42.1 on Ollama for GPT-OSS-120B.
- [Every GB10 machine](https://hasso5703.github.io/machines/)
- [NVIDIA DGX Spark](https://hasso5703.github.io/machines/nvidia-dgx-spark/): 128 GB from NVIDIA. Since October 23, 2026, NVIDIA's partners also sell 64 GB configurations.
- [ASUS Ascent GX10](https://hasso5703.github.io/machines/asus-ascent-gx10/): Sold with 128 GB, and with 64 GB since October 2026.
- [Dell Pro Precision with GB10](https://hasso5703.github.io/machines/dell-pro-max-gb10/): Launched as the Dell Pro Max with GB10, now named Dell Pro Precision with GB10 on dell.com, configured with 64 GB or 128 GB.
- [HP ZGX Nano G1n AI Station](https://hasso5703.github.io/machines/hp-zgx-nano-g1n/): 128 GB on HP's US page. NVIDIA lists HP among the partners selling 64 GB configurations.
- [Lenovo ThinkStation PGX](https://hasso5703.github.io/machines/lenovo-thinkstation-pgx/): 128 GB.
- [MSI EdgeXpert](https://hasso5703.github.io/machines/msi-edgexpert/): Sold with 128 GB, and with 64 GB since October 23, 2026.
- [GIGABYTE AI TOP ATOM](https://hasso5703.github.io/machines/gigabyte-ai-top-atom/): Sold with 128 GB, and with 64 GB since October 2026.
- [Acer Veriton GN100 AI Mini Workstation](https://hasso5703.github.io/machines/acer-veriton-gn100/): Sold with 128 GB or 64 GB.

## Source
- [Repository](https://github.com/hasso5703/dgx-spark-qwen38)
- [Benchmarks and methods](https://github.com/hasso5703/dgx-spark-qwen38/blob/main/BENCHMARKS.md)
- [Clients: Claude Code, opencode, Open WebUI, Cursor, VS Code Copilot](https://github.com/hasso5703/dgx-spark-qwen38/blob/main/docs/clients.md)
- [The GB10 unified-memory trap](https://github.com/hasso5703/dgx-spark-qwen38/blob/main/docs/gb10-memory.md)
- [Releases](https://github.com/hasso5703/dgx-spark-qwen38/releases)
- [NVIDIA forum thread](https://forums.developer.nvidia.com/t/380257)