Open Weights Week: DeepSeek Publishes 284B MIT Model; Anthropic Draws Its Line on What It Won't Restrict

DeepSeek released a 284B MIT-licensed MoE model on July 31; Anthropic published a position opposing blanket bans on open-weights models while backing safety-test mandates and distillation controls.

DeepSeek V4 Flash 0731

The July 31 build is a production-candidate refresh of DeepSeek's V4 Flash architecture: 284B total parameters, 13B active per forward pass (Mixture of Experts), MIT license, available on Hugging Face. The API runs at ~122 tokens/second.

Practical caveat for self-hosters: "Flash" refers to the active-parameter efficiency, not the weight footprint. Full weights are 140GB+ in mixed precision, 70-80GB heavily quantized — you need multi-GPU server hardware to run it locally. For single-GPU or CPU inference, smaller open models remain more practical.

Anthropic's open-weights stance

In a July 27 post, Anthropic stated it has never advocated for a blanket ban on open-weights models. What it does support: mandatory safety evaluations for all sufficiently capable models (open and closed), cracking down on industrial-scale distillation, and restricting advanced chip exports — but not open-weights releases categorically.

Why it matters: V4 Flash weights are commercially usable under MIT. On policy, the major lab position is now explicit: chip controls and distillation restrictions are the levers, not a blanket ban on open release.

Sources: DeepSeek-V4-Flash — HuggingFace · V4 Flash Analysis — Artificial Analysis · Anthropic Open Weights Position