model · model/deepseek-deepseek-v4-flash-0731

DeepSeek: DeepSeek V4 Flash 0731

Also called DeepSeek-V4-Flash-0731, deepseek/deepseek-v4-flash-0731

Facts

price input
$0.06per million tokensopenrouter-models, last checked 2026-09-14
price output
$0.12per million tokensopenrouter-models, last checked 2026-09-14
context window
1310720tokensopenrouter-models, last checked 2026-09-14
intelligence index
34.5openrouter-models, last checked 2026-09-14
coding index
69.1openrouter-models, last checked 2026-09-14
status
activeopenrouter-models, last checked 2026-09-14
license
MIT License, repository and weightssource, accessed 2026-08-29
parameters
284B total, 13B active per token (sparse mixture of experts)source, accessed 2026-08-29
repository tensor total
304B paramssource, accessed 2026-08-29
preview parameters
284 billion parameterssource, accessed 2026-08-29
structure
the same model structure as DeepSeek-V4-Flash-DSpark, i.e. it comes with a speculative decoding module attachedsource, accessed 2026-08-29
terminal bench score
82.7 on Terminal Bench (2.1), per the model cardsource, accessed 2026-08-29
cybergym score
76.7 on CyberGym, per the model cardsource, accessed 2026-08-29
preview terminal bench score
61.8 on Terminal Bench (2.1) — the April preview checkpoint's score, per the same table on the July model cardsource, accessed 2026-08-29
preview cybergym score
38.7 on CyberGym — the April preview checkpoint's score, per the same table on the July model cardsource, accessed 2026-08-29
release date
2026-07-31source, accessed 2026-08-29

Timeline

  1. released with MIT-licensed weights on Hugging Facesource
  2. V4 series previewed, including the 284-billion-parameter DeepSeek-V4-Flash that this row's release supersedessource