.svg){kind=spacer}
Details
- DeepSeek released DeepSeek-V4 Preview, fully open-sourced, featuring cost-effective 1M context length as the new standard.
- DeepSeek-V4-Pro has 1.6T total parameters and 49B active parameters, rivaling top closed-source models in performance.
- DeepSeek-V4-Flash has 284B total parameters and 13B active parameters, offering reasoning close to V4-Pro with faster responses and lower API costs.
- Key innovations include novel token-wise compression and DeepSeek Sparse Attention (DSA) for ultra-high context efficiency, reducing compute and memory costs.
- Enhanced agentic capabilities lead open-source models in Agentic Coding benchmarks; rich world knowledge trails only Gemini-3.1-Pro; excels in Math/STEM/Coding.
- API available immediately via existing base\_url with models deepseek-v4-pro or deepseek-v4-flash, supporting OpenAI ChatCompletions, Anthropic APIs, 1M context, and dual Thinking/Non-Thinking modes.
- Integrated with AI agents like Claude Code, OpenClaw, and OpenCode; powers DeepSeek's in-house agentic coding.
Impact
DeepSeek-V4's open-sourcing of 1M context models at low cost pressures proprietary leaders like OpenAI and Anthropic, who charge premiums for similar lengths, accelerating adoption in agentic workflows and long-document tasks. V4-Pro's benchmark leadership among open models narrows the gap with closed rivals, while Flash's efficiency enables broader enterprise use without sacrificing reasoning. This advances cost-effective AGI trajectories, potentially shifting market dynamics toward accessible high-context inference amid rising compute demands.
