What happened
DeepSeek released the final V4-Flash checkpoint and opened its official API in public beta. The 284-billion-parameter mixture-of-experts model activates approximately 13 billion parameters, supports a one-million-token context window, and is available as downloadable weights under an MIT license. Its hosted service supports OpenAI-, Anthropic-, and Responses-compatible interfaces.
Potential direction of power
Dispersion · Dispersion +1
The release combines independently deployable weights, unusually low inference prices, and interface compatibility that lowers the cost of substituting DeepSeek for incumbent providers. This creates meaningful separation between the model and DeepSeek’s hosted service, although practical independent deployment still requires substantial hardware and technical capacity.
Window of contestability
Opens
The weights, license, model card, deployment materials, and interfaces are publicly accessible. Training data and full development provenance remain unavailable, while practical contestability will depend on independent hosting, ecosystem support, and demonstrated provider substitution.
