> For the complete documentation index, see [llms.txt](https://docs.nvcl.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.nvcl.ai/benchmarks/other-models-and-hardware.md).

# Other models

Results beyond Qwen3.8-27B: DeepSeek-V4.1-Flash on two H200.

## DeepSeek-V4.1-Flash on two H200

NVCL AI runs the whole model on two H200 plus server memory, where vLLM's published recipe for this model on H200 needs 4 cards and recommends 8. It serves DeepSeek's published weights, unchanged.

| Chats       | Drafts on, greedy, tok/s | Drafts on, sampled (T 1.0), tok/s | Drafts off, greedy, tok/s |
| ----------- | -----------------------: | --------------------------------: | ------------------------: |
| 1           |                      302 |                               300 |                       165 |
| 2, in total |                      399 |                               412 |                       284 |

The model's built-in drafts (speculative decoding) give 1.83x the output of drafts off for one chat and about 1.4x for two, and in these runs greedy answers were the same token for token with drafts on and off. The first token of these short prompts comes in about 0.06 s.

<details>

<summary>Conditions</summary>

* DeepSeek-V4.1-Flash, checkpoint Flash-1017.
* 2x H200 plus about 240 GB of host memory (the engine used about 232 GB at its peak).
* Chat defaults with the model's built-in drafts, and the same runs with drafts off; 6 short prompts, 512-token answers; three runs per setting, each from a fresh load.
* NVCL AI build of 2026-10-23. Measured 2026-10-23.

</details>