AI BriefingDeepSeek-V4 Series Unveiled
Key point
A DeepSeek-V4 preview has been released, supporting a 1.6T MoE and 1M context.
Details
DeepSeek has released a preview version of DeepSeek-V4 on Hugging Face.
The series consists of two models.
- DeepSeek-V4-Pro: 1.6T parameters, 49B activated, 1M token context
- DeepSeek-V4-Flash: 284B parameters, 13B activated, 1M token context
The core features are Hybrid Attention and mHC. The company stated that at the 1M token range, single-token inference FLOPs were reduced to 27% compared to DeepSeek-V3.2, and KV cache was compressed to around 10%.
It also explained that the Muon Optimizer was applied to improve training stability and convergence speed, and the instruct model supports three reasoning modes: Non-think / Think / Think Max. Code, math, and long-context benchmark results were also released together.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Jevper Released: A Jev Classification Wrapper for OpenAI-Compatible Models with Automatic Logprobs Detection
Rootless Linux Sandbox 'Drop' Released for Isolating AI Agents and Third-Party Programs