DeepSeek Launches V4 Preview: 1M Context, Open-Source, Two Model Tiers
DeepSeek open-sources V4-Pro and V4-Flash with 1M-token context windows; API available today.
DeepSeek has released and open-sourced DeepSeek-V4 Preview, introducing two models: DeepSeek-V4-Pro (1.6 trillion total parameters, 49 billion active) and DeepSeek-V4-Flash (284 billion total, 13 billion active). Both support 1M-token context windows by default and offer dual Thinking/Non-Thinking modes. API access is live today — developers switch by updating the model name while keeping existing base URLs, with support for both OpenAI ChatCompletions and Anthropic APIs.
For builders and buyers, the release pushes 1M-token context into the open-source mainstream at reduced compute and memory cost, enabled by a new architecture combining token-wise compression and DeepSeek Sparse Attention. V4-Pro claims open-source state-of-the-art on agentic coding benchmarks, trailing only Gemini-3.1-Pro on world knowledge while beating all current open models in math, STEM, and coding. DeepSeek says both models are already integrated with Claude Code and other coding agents, with V4-Flash delivering near-V4-Pro reasoning at smaller parameter counts and lower cost. The announcement came via DeepSeek's official X account.