Two models, and why exactly these two
Every machine we build runs one of two open models, and we mod both. Both run on NVIDIA hardware only, in the Blackwell-optimised NVFP4 build — 4-bit in the tensor cores' native format, the configuration we ship, test and mod.
The smartest model that fits in a desk-side box. Long-memory roles.
The smartest model that fits in a desk-side box, and what the long-memory roles run. The model card gives 125B of language with 6B active per token, plus a 51B n-gram embedding table and a 4B MTP head: 176B of weights, 180B packaged, which is the size Hugging Face reports for the checkpoint. It moves like a small model while knowing what a large one knows — 43.0 tok/s on a DGX Spark generating text, flat from 327 to 258,790 prompt tokens, though the same box measures 88.5 tok/s reproducing a file and 27.8 on free-form prose, which is why the task shape is named with every figure here — sees text, images and video, and its cache costs about 13 KB per token in the Q8 configuration we ship, 24 KB at full precision, DERIVED FROM THE ARCHITECTURE with no vendor figure published, which is what makes its full 262k window practical on one box.
- Configuration: NVFP4 4-bit, Q8 cache
- Context: 262k native
- Memory: 123.0 GB
- Runs on: NVIDIA DGX Spark (GB10), 4U4G-TURIN/HPR, NVIDIA DGX Station (GB300), 4UXGM-TURIN2 DIRECT
Cheap, strong text and code at a 1M window — and it reads images.
The value pick: near its bigger sibling's quality for a fraction of the memory, under the most permissive licence in the catalogue, and a native 1M-token window with a cache cheap enough to actually use it. It activates 8B parameters per token while reading the prompt and 16B while writing the answer — the prefill figure is the one that matters for the long agentic work it is bought for. It reads documents itself: DeepSeek's own vision encoder, measured on its own table at 95.6 on DocVQA and 86.0 on RefCOCO. DeepSeek's figures for it are on the page below, each with the card it came from.
- Configuration: NVFP4 4-bit, Q8 cache
- Context: 1M native
- Memory: 506.3 GB
- Runs on: 4U4G-TURIN/HPR, NVIDIA DGX Station (GB300), 4UXGM-TURIN2 DIRECT, 8U16X-TURIN2 B300
Why exactly these two
One model that sees: Qwen3.8-Flash-Next reads invoices, drawings and site photos, holds a long memory, and does it in a desk-side box. One model that thinks cheaply and at length: DeepSeek V4.1-Flash carries a native 1M-token window under MIT terms for text-heavy work at length, and its own vision encoder reads a document when one turns up. Where a job leans on both, they run side by side.
Both run on NVIDIA hardware only. That is the configuration we ship, test and mod; we do not sell or support other accelerators, because the thing being bought is a working setup, not a parts list.
Want to know what fits the machine you already own?
Tell us the model you are weighing and the hardware you have, and we will say whether it fits, what the configuration would be and what it costs in memory — before any conversation about price.
Email us · Register for a role