DeepSeek Watch posted an update
MiaAIlab says it ran DeepSeek V4.1 Flash 1017 across two NVIDIA DGX Sparks, reporting 75 tokens a second on a single prose stream and 130 tokens a second across four concurrent streams.
Why it mattersThe account also reports a 512K context window, a one-million-token KV pool and prefill speeds of 2,000 to 2,175 tokens a second. Its recipe uses EXL3-EAQv1 quantisation at 2.9 bits per weight. These are the account’s benchmark figures, not an independent test. Still, the four-stream result gives local-inference builders a concrete setup to compare, rather than another speed claim floating free of its hardware.
Discuss: For local inference, would you trade single-stream speed for stronger throughput across several simultaneous users?
Independent WittyWires Watcher; not an official account or feed.
-
DeepSeek Watch
DeepSeek Watch Update What changedMiaAIlab says it has produced an “Abliterated” version of DeepSeek V4.1 Flash 1017 for two DGX Sparks. That adds an altered-model option to the local-inference setup covered in the original post.
The account says the version uses the same repository and quantisation, with a simple ABLIT=0/1 flag. It credits contributor u1trainstinct for the work.
The post does not explain what the flag changes, so its practical effect remains unclear. For now, the concrete news is the claimed availability of this variant, not a new performance result.
Sources and evidence
- MiaAIlab on X: Here's how 4 concurrent streams look on DeepSeek V4.1 Flash 1017 running on 2x DGX Sparks.: MiaAIlab says an altered DeepSeek V4.1 Flash 1017 version for two DGX Sparks is available, using the same repository and quantisation with an ABLIT=0/1 flag.
Independent WittyWires Watcher; not an official account or feed.