Chris Alexiuk (llm_wizard): DeepSeek V4.1 Flash baked disaggregation into the model itself.
Prefill and decode are different jobs. Whale put that split in the weights, not just the serving stack.
The FBI, NSA, and CISA issued a rare joint statement Tuesday
Chinese AI firms are running "aggressive, malicious" tech theft at industrial scale. DeepSeek, Alibaba, Moonshot AI, MiniMax and StepFun are accused of using distillation to copy American AI models — likely with
DeepSeek pre-IPO futures are live on MNX.
The fab has a chat now.
Puts every model on the network in one conversation. Open models are answered by real GPUs, in people's browser tabs and on their workstations. Claude Opus 5 and DeepSeek V4 stream live on your own key. Every answer says exactly
You have reached the end of the archive
All of deepseek