dgx-spark-qwen38 / Machines / Lenovo ThinkStation PGX
Local AI on the Lenovo ThinkStation PGX
Also known as ThinkStation PGX, Lenovo PGX. 128 GB.
The Lenovo ThinkStation PGX is a GB10 machine: NVIDIA's Grace Blackwell chip and its unified memory, on DGX OS. One command turns it into a private AI server with OpenAI- and Anthropic-compatible APIs and a local cockpit, and no prompt, image or video leaves the box.
The project is measured on an ASUS Ascent GX10, the same chip and the same system. The installer reads the memory of the box, not its brand. If you run it on a Lenovo ThinkStation PGX, tell us how it went in the discussions.
With 128 GB: every lane
| Lane | Measured |
|---|---|
| Qwen3.8-27B | 71.4 tok/s, greedy median, one request |
| Qwen3.8-Flash-Next 176B | 47.9 tok/s, on code, with FLASH_TIER=context |
| Flash FP4-Local | 55.4 tok/s, greedy median, at its default effort |
| Nemotron 3.5 Bolt | 140 tok/s, rewriting code, one request, reasoning off |
| Nemotron 3.5 Super VL | 69.5 tok/s, rewriting code, one request, reasoning off |
| Ornith 1.5 122B | 42.4 tok/s, greedy median, one request |
| Kolibri-1 | 48.0 tok/s, one request |
| Qwen-Image 2.1 | 7.5 s, with its Turbo, per 1024×1024 image |
| MiniMax-H3 | 126.2 to 127.9 s, with the Turbo adapter, for 4 s of 480P video |
Install it
On stock DGX OS, run this on the box. The same command updates it later.
curl -fsSL https://raw.githubusercontent.com/hasso5703/dgx-spark-qwen38/main/get.sh | bash- Numbers
- measured on an ASUS Ascent GX10 with 128 GB, methods in BENCHMARKS.md
- Maker's page
- Lenovo ThinkStation PGX