DeepSeek 开源推理引擎 DeepInfer：支撑 V4.1 Flash，Apache 2.0
DeepInfer 以 Apache 2.0 开源，支撑 $0.30/$1.20 每百万 token 的 V4.1 Flash，用投机解码与专家分布调度，仅支持 Blackwell 和昇腾 950，未含 KV cache 与 PD 分离。
SITUATION EXPLAINED: DeepSeek just open-sourced the engine that runs its AI, something OpenAI and Anthropic haven't done. • DeepInfer powers DeepSeek V4.1 Flash, which costs $0.30 per million input tokens and $1.20 per million output, and it's now open source under Apache 2.0 • Inference is running a trained model, and serving a 552B-parameter one fast and cheap to millions of users is one of AI's hardest engineering problems • It spreads users, long prompts, and expert sub-networks across GPUs, and uses speculative decoding, where a small model guesses ahead and the big one checks • It only runs on Nvidia Blackwell GPUs and Huawei Ascend 950 chips, so it's built for data centers, not home rigs • It still leaves out the KV cache system and prefill/decode disaggregation, so it's not a full serving stack @theojaffee: "DeepSeek open sources almost everything. It's really incredible. They release their model weights. They write papers explaining in detail their model architecture. They put out GitHub repos with their inference framework. DeepSeek really cares about fully open AGI in a way that other companies don't."
这条是引用着另一条写的：「DeepSeek 开源推理框架 DeepInfer，支撑 V4.1 Flash 生产服务」