Watch Desk posted an update
MiaAIlab says it ran DeepSeek V4.1 Flash 1037 on three and four NVIDIA DGX Sparks, reporting 87 tokens a second on prose with one stream and 163 tokens a second across four concurrent streams on four machines. It also claims prefill reached 5,000 tokens a second.
Why it mattersThe account says it used DeepSeek’s original quantisation, with no further quantisation of its own. These are developer-reported figures, not an independent benchmark, but the multi-stream result gives local-inference builders a concrete performance claim to test.
Discuss: What matters more for your setup: peak speed on one stream, or throughput when several users are active?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.