DeepSeek · Model release
DeepSeek V4-Flash-0731: why the post-training update matters
Retired from the first-party API on 2026-09-10: DeepSeek says the legacy `deepseek-v4-flash` id temporarily routes to V4.1-Flash at its rates. This page preserves a historical note about the 0731 revision; it is not a current model recommendation.
What changed
- DeepSeek's first-party API documentation says the `deepseek-v4-flash` model was updated to V4-Flash-0731; the call identifier remained unchanged at that time.
- A July 31 secondary report described post-training and serving changes, but the primary-source record reviewed here does not establish the full architecture or benchmark methodology; treat those historical details as unverified.
- Do not use this historical page for current model limits or rates; consult the current V4.1-Flash model and pricing documentation.
Best fit
- Reviewing July 2026 release history
- Tracing the API ID transition
- Historical comparison only
What to be careful about
- Any benchmark numbers in the July 31 secondary report are not independently verified here and may use a DeepSeek harness.
- Open weights can still require datacentre-class memory.
- Retired on the first-party API since 2026-09-10 — consult the current DeepSeek V4.1-Flash API and rate table for any new work.
Dataset snapshot: DeepSeek V4 Flash is also represented in the checked comparison dataset. Prices and benchmark figures carry their own field dates.