AI Radar

DeepSeek V4 Pro 0813 appears on Hugging Face

A Hacker News post links a Hugging Face page for DeepSeek-V4-Pro-0813, described as the official release superseding the preview. The page claims better agentic ability and adds a DSpark speculative decoding module.

Who should care
Developers using DeepSeek models for agent and coding work.

Why it matters: The MIT licence would allow local and commercial use of the weights if the listing is confirmed.

Key facts

License
MIT
Terminal Bench 2.1
87.9 vs 72.1 for the preview
DeepSWE
62.7 vs 12.8 for the preview
Cybergym
83.3 vs 52.7 for the preview
Page compares against Kimi K3, GLM-5.2 and Opus-4.8

Context

The comparison lists Kimi K3 at 88.3 on Terminal Bench 2.1, slightly above the DeepSeek score.

Timeline

Discussion

  1. DeepSeek V4 Pro 0813huggingface.co · HN · 25 pts · 1 comments

Evidence

Low evidence: a single community post. Treat it as a claim until someone else confirms it.

Details written Nov 14, 11:38 UTC from the sources listed above.