I feel like we gotta push the extremes of one idea to get to the fringes.

So let's fit 32 deepseek_a users on 2 sparks. TP2 v0.7.5 on 2× GB10: -32 users at once (was 16) -249 tok/s total (+25%), ~8 tok/s each -

Bertholomus AINov 48

This is even more insane, DeepSeek-v4.1-flash build me a complete game in one-shot attempt.

Not with any harness, not using any external assets, no tool calls, just single turn, one prompt in and html output. Model is DeepSeek-v4.1-Flash-1017 (deepseek-flash is the model name

AJNov 41

DeepSeek

EdgeTypENov 4

DeepSeek open-sourced DeepInfer, the engine behind its API.

Its receipt: $38 of revenue per GPU-hour (off-peak). In 2025 the same math gave ~$13 on H800s. The catch: 64 + 64 Blackwell-class GPUs, and no cache store in the repo.

Seb MoskaNov 4