Nicholai Mitchko releases self-hosted DeepSeek latent-reasoning stack for Blackwell GPUs
The InterSystems AI director paired a CoLaR reasoning head with a DeepSeek-V4-Flash backbone whose routed MoE experts use NVFP4, while attention, shared experts, the LM head and draft block remain at higher precision. Its sole published benchmark has conflicting aggregate figures and no independent replication.
By Ryan Merket · Published
Primary source: Nicholai Mitchko's AI Research blog
Why it matters
Mitchko has made a latent-reasoning system available as weights plus runnable serving code, giving inference teams a concrete test of hidden-state computation. Its specialized vLLM path, heavy Blackwell memory requirement and unreplicated benchmark show how far the approach remains from routine deployment.
Nicholai Mitchko, InterSystems' director of AI enablement, 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.