The model, trained with a vastly improved post-workout pipeline, now exceeds the V4-Pro preview on all agentic benchmarks published by DeepSeek.
DeepSeek has made official DeepSeek-V4-Flash-0731, the production version of its Flash model, released as open source under the MIT license on July 31, 2026. It is an expert mix (MoE) model of 284 billion parameters, including 13 billion active, designed for code, tool usage, and agentic workflows, with a context window of one million tokens and up to 384,000 output tokens.
According to DeepSeek, the architecture remains the same as the April preview: performance gains come entirely from a largely reworked post-workout rather than an increase in model size. The announced pricing is $0.14 and $0.28 per million tokens, with a 98% discount on cached tokens.
Read the full article on the original source
Go to the source →