A mixture-of-experts model with 284B total parameters and only 13B active per token, a one-million-token window, and weights published with no access gating. The agentic gains are large: 82.7 vs 72.1 on Terminal Bench 2.1 and 54.4 vs 12.8 on DeepSWE.
Released on July 31, DeepSeek-V4-Flash-0731 is a mixture-of-experts (MoE) architecture with 284 billion total parameters and 13 billion active per token, a one-million-token context window, one shared expert and 256 routed experts of which six fire per token. The first three MoE layers use hash routing.
The agentic results (versus the previous version): Terminal Bench 2.1, 82.7 vs 72.1; NL2Repo, 54.2 vs 38.5; Cybergym, 76.7 vs 52.7; DeepSWE, 54.4 vs 12.8.
The pricing: $0.14 per million input tokens on a cache miss, $0.0028 on a cache hit, and $0.28 per million output tokens.
Why it matters: the weights are published under an MIT licence with no access gating. Just as Europe starts demanding transparency about training data and Washington debates what to do about Chinese open models, a model of this size appears that anyone can download and run. The open-weights policy debate stops being theoretical.
Go to the full analysis →Source: MarkTechPost · Hugging Face
Go to the original source →