DeepSeek launches DeepSeek-V4-Pro and V4-Flash preview

DeepSeek released a preview of two new open-weight models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, publishing weights under the permissive MIT licence and making both available through its API on the same day. Pro is a mixture-of-experts model with 1.6 trillion total parameters and 49 billion active per token; Flash is a smaller variant with 284 billion total and 13 billion active parameters, built for faster and cheaper inference. Both support a context window of one million tokens and combine “thinking” and “non-thinking” modes in a single model, using a sparse attention mechanism DeepSeek said was designed specifically to make long-context processing efficient.

DeepSeek said Pro led open-weight competitors on agentic coding benchmarks and on general world-knowledge measures trailed only Google’s Gemini 3.1 Pro among all models, closed or open, while matching leading closed models on reasoning tasks. Flash was pitched as retaining reasoning performance “closely approaching” Pro’s, at substantially lower cost — DeepSeek priced the two models well below the leading closed-weight alternatives on a per-token basis.

The release continued a pattern established with DeepSeek’s earlier V3 and R1 models: publishing frontier-class capability as freely licensed weights rather than behind an API-only wall, which by 2026 had made DeepSeek one of the few labs still contesting the top of open-weight leaderboards against a shrinking field of Western competitors, several of which — including Meta with Muse Spark, released the same month — had moved toward closed release. The MIT licence meant the models could be self-hosted and modified commercially without contacting DeepSeek, extending the company’s practice of making frontier weights available to anyone with the hardware to run them.