DeepSeek-V4-Flash

RETIRED 2026-09-10, superseded by DeepSeek-V4.1-Flash (deepseek-flash). Signal: an explicit past-tense statement by the provider on a dated announcement — DeepSeek's Change Log entry for 2026-09-10 reads 'The previous-generation models V4 Flash and V4 Flash Vision Exp have been retired; for compatibility, the model names deepseek-v4-flash and deepseek-v4-flash-vision-exp are temporarily routed to V4.1 Flash', and the pricing page footnote says the same: 'The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but the corresponding models have been retired, their requests are served by the DeepSeek-V4.1-Flash model and billed at the Flash price.' That is a statement, not an inference from the column vanishing off the pricing page, though it vanished the same day. DeepSeek published no separate shutdown date and no deprecation window: one date, so this row carries it in retires_on only and leaves deprecated_on null. THE NAME STILL WORKS — calls do not fail, they are served by V4.1-Flash and billed at the cheaper Flash price, and DeepSeek calls that routing 'temporary', so consumers should migrate to deepseek-flash rather than rely on it. Until 2026-09-10 this row was carried as preview, on the grounds that it was part of the 'DeepSeek V4 Preview' generation released 2026-04-24, and that was right for the period it covered. PRICES ABOVE ARE THE RATES IN FORCE AT RETIREMENT and deliberately no longer track the endpoint that now serves the traffic: PEAK $0.014/M cache-hit, $0.44/M cache-miss, $1.32/M out, with off-peak exactly half. V4.1-Flash is cheaper on every axis ($0.006 / $0.30 / $1.20 peak), so a caller left on the legacy name is billed at the new, lower Flash price, not at these. Historical spec, unchanged: smaller/cheaper V4 model (~284B total / ~13B active params per authoritative third-party reports), 1M context, 384K max output, dual Thinking / Non-Thinking modes, JSON output, tool calls, chat prefix completion, FIM completion in non-thinking mode only, Responses API and Anthropic-format API supported, concurrency limit 2500, open weights. Knowledge cutoff never published by DeepSeek, so it stays null. The legacy aliases deepseek-chat and deepseek-reasoner pointed here and were themselves retired 2026-07-24, which makes that pair a two-hop chain to deepseek-v4-1-flash; each row records the target its provider named at the time rather than being silently re-pointed. Prior lifecycle of this row, kept because it explains the price fields: DeepSeek moved API billing to peak/off-peak at 16:00 UTC on 2026-08-16, having announced it undated on 2026-08-13 and dated on 2026-08-14; the fields were re-pointed at the PEAK column on the cutover day, about 8 hours early, because this catalog refreshes once daily at ~08:05 UTC and the alternative was publishing superseded rates for ~16 hours. That window closed on 2026-08-17.

Source: DeepSeek official documentation ↗

Explore more

DeepSeek-V4-Flash — questions & answers