DeepSeek releases V4-Pro and V4-Flash: open-source MoE models with 1M-token context and sparse attention for a fraction of frontier costs

DeepSeek on April 24, 2026 released DeepSeek-V4-Pro and DeepSeek-V4-Flash simultaneously — two open-source Mixture-of-Experts models that bring 1-million-token context windows, a new sparse attention architecture, and frontier-competitive performance to the API at a fraction of the cost of closed-source alternatives.

What's new

DeepSeek-V4-Pro is a 1.6-trillion-parameter MoE model with 49 billion parameters active per forward pass. DeepSeek describes it as "Performance rivaling the world's top closed-source models" on math, STEM, and coding benchmarks, with rich world knowledge trailing only Gemini 3.1 Pro among open models.

DeepSeek-V4-Flash is the smaller sibling: 284 billion total parameters, 13 billion active, designed for fast and cost-effective inference with reasoning capabilities approaching V4-Pro on agent tasks.

Both models:

• 1M-token context window — now the standard for all DeepSeek API offerings
• Token-wise compression + DeepSeek Sparse Attention (DSA) — a new architecture that delivers "world-leading long context with drastically reduced compute & memory costs"
• Support both OpenAI ChatCompletions and Anthropic API formats, minimizing migration effort
• Native integration with Claude Code, OpenClaw, and OpenCode agent frameworks
• Both models are fully open-sourced

Legacy models deepseek-chat and deepseek-reasoner are scheduled to retire on July 24, 2026; users should migrate to V4-Pro or V4-Flash before that date.

Context

DeepSeek has become a reference point for cost-efficient frontier AI. DeepSeek-V3 set a benchmark when it was open-sourced in late 2025, demonstrating that capable MoE models could be built at dramatically lower training cost than US frontier labs. V4 continues that trajectory: 1M-token context is now standard across both the lean Flash variant and the large Pro model.

The DSA (DeepSeek Sparse Attention) mechanism is the key architectural advance in V4. Standard attention scales quadratically with sequence length; sparse attention approaches — including DeepSeek's token-wise compression — reduce that cost, making long-context inference commercially viable at the pricing DeepSeek offers.

Why it matters

DeepSeek V4 strengthens the open-source case for production AI deployment. V4-Pro's benchmark parity with closed frontier models, combined with the standard API compatibility and open weights, means teams can self-host or use the DeepSeek API without compromising capability relative to OpenAI or Anthropic offerings.

The 1M-token context window at DeepSeek's price points undercuts what US providers charge for the same context length. For workloads like codebase analysis, legal document review, or long research sessions, the economics shift significantly toward open alternatives.

The Huawei Ascend NPU support (noted in the launch materials) also signals DeepSeek's intent to provide a viable compute path outside NVIDIA's H100/H200 ecosystem — a notable detail given US export controls on advanced chips to China.