Skip to main content
AI Choice Engine
← All launch guides

DeepSeek · Model release

DeepSeek V4.1-Flash: the hosted successor to retired V4-Flash

DeepSeek's September release is an MIT-licensed multimodal model. DeepSeek reports comparisons against V4-Pro; those vendor results are not an independently matched ranking. The same announcement retired V4-Flash API aliases.

DeepSeek

Released September 10, 2026

Compare it

What changed

  • The hosted September 10 launch uses `deepseek-flash` with a 1M context limit, 384K max output and image input. The separately downloadable MIT checkpoint has a native 1M-context publisher label; the API output cap is not a verified native generation limit.
  • V4-Flash and V4-Flash-Vision-Exp were retired from the API; their legacy ids temporarily route to V4.1-Flash and are billed at its current rates.
  • The current price page lists $0.30/$1.20 per MTok at peak and $0.15/$0.60 off-peak, with separate cache-hit rates.

Best fit

  • Cost-sensitive agents
  • Large-scale open-weight evaluation
  • High-volume coding

What to be careful about

  • Peak hours are 01:00–04:00 and 06:00–10:00 UTC Monday through Friday, excluding Chinese public holidays; weekends and those holidays are off-peak.
  • DeepSeek reports Terminal-Bench 2.1 and DeepSWE 1.1 results at max effort on its model card; these are vendor-run and not independently rerun here.
  • The native card specifies a 552B backbone plus 196B of Engram conditional memory, with 8B active during prefill and 16B during decode. The published minimal inference reference is not a production serving engine. Mixed precision, lookup-memory placement, cache and backend support prevent a uniform 552B Q4 calculation from proving desktop fit.
Dataset snapshot: DeepSeek V4.1 Flash is also represented in the checked comparison dataset. Prices and benchmark figures carry their own field dates.