Nicholai Mitchko 发布了面向 Blackwell GPUs 的自托管 DeepSeek 潜在推理堆栈
InterSystems 的 AI 总监将一个 CoLaR 推理头与 DeepSeek-V4-Flash 主干配对，其路由的 MoE 专家使用 NVFP4，而注意力、共享专家、LM head 和 draft block 则保持更高精度。其唯一公布的基准测试在汇总数据上存在冲突，且没有独立复现。
By Ryan Merket · Published
Primary source: Nicholai Mitchko's AI Research blog
为何重要
Mitchko 已将一个潜在推理系统以权重加可运行的服务代码形式提供，给推理团队提供了对隐藏状态计算的具体测试。其针对 vLLM 的专用路径、对 Blackwell 的高内存需求以及未被复现的基准，显示出该方法距离常规部署还有很大差距。
Nicholai Mitchko, InterSystems 的 AI 赋能总监, has released a self-hosted latent-reasoning model built around DeepSeek-V4-Flash. The model performs intermediate computation in hidden states before generating a visible answer, but serving it requires Blackwell-class Nvidia hardware and Mitchko's specialized inference code.…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.