From V4 Flash to V4.1 Flash
DeepSeek V4.1 Flash is not a retune of V4 Flash. It is the first model of a new architecture family, released on 10 September 2026, and DeepSeek retired V4 Flash the same day.
The timeline
| Date | Release | What DeepSeek said |
|---|---|---|
| 24 Apr 2026 | V4 Pro and V4 Flash (preview) | Both on the API in OpenAI and Anthropic formats; the old deepseek-chat and deepseek-reasoner names map to V4 Flash until 24 July 2026. |
| 31 Jul 2026 | V4 Flash 0731 | Public beta of the V4 Flash API, same architecture and size as the preview, with a native Responses API adapted for Codex. |
| 10 Sep 2026 | V4.1 Flash | The smallest model of a new architecture family, with native visual understanding. V4 Flash and V4 Flash Vision Exp retired. |
| 17 Oct 2026 | V4.1 Flash 1017 | A minor update: better agent and coding scores, shorter outputs, same price. |
From DeepSeek's changelog and 10 September announcement.
What is different inside
V4.1 Flash is a 552B mixture-of-experts model that activates 8B parameters per token while reading the prompt and 16B while writing, on what DeepSeek calls a new causal encoder-decoder architecture. DeepSeek says it needs a quarter of the GPU memory and an eighth of the SSD storage for its KV cache compared with the previous generation, which is what makes a million-token context cheap to serve. It also reads images natively, where V4 Flash needed a separate vision variant.
On the API, deepseek-flash is the name for the new model. DeepSeek routes the old deepseek-v4-flash and deepseek-v4-flash-vision-exp names to V4.1 Flash for now, so a config that still names V4 Flash is already running V4.1.
DeepSeek's own numbers, release against release
| Benchmark | V4 Flash 0731 | V4.1 Flash |
|---|---|---|
| Terminal Bench 2.1 | 82.7 | 90.6 |
| NL2Repo | 54.2 | 65.4 |
| DeepSWE | 54.4 | 74.2 |
| Agents' Last Exam | 25.2 | 31.8 |
Each column is from that release's changelog entry. DeepSeek names DeepSWE "v1.1" only in the V4.1 entry, so treat the row as indicative rather than a like-for-like rerun.
Artificial Analysis
| Model | Intelligence Index |
|---|---|
| DeepSeek V4.1 Flash (max effort) | 39 |
| DeepSeek V4.1 Flash (no reasoning) | 25 |
| DeepSeek V4 Flash 0731 (max effort) | 34 |
Intelligence Index v4.3.2, checked 18 October 2026. Artificial Analysis lists V4 Flash as deprecated.
Five points is a large step for a same-name release, and most of it comes from reasoning: the same model with reasoning off scores about 25.
Running either one uncensored
Wild West API sells DeepSeek V4.1 Flash Outlaw (deepseek-v4.1-flash-outlaw): an uncensored DeepSeek V4.1 Flash, text only, with tool calling and reasoning on by default. There is no uncensored V4 Flash on the line, and with V4 Flash retired by DeepSeek there is no reason to want one. The API quickstart has the exact request.
FAQ
Is DeepSeek V4.1 Flash better than V4 Flash?
On every benchmark DeepSeek published for both, yes, and Artificial Analysis scores it 39 against 34 for V4 Flash 0731 at maximum reasoning effort.
Is V4 Flash still available on DeepSeek's API?
DeepSeek retired it on 10 September 2026. The deepseek-v4-flash name is temporarily routed to V4.1 Flash.
Does V4.1 Flash read images?
DeepSeek's own release does, natively. The uncensored DeepSeek V4.1 Flash Outlaw on Wild West API is text only.
How big is DeepSeek V4.1 Flash?
DeepSeek describes it as a 552B-parameter mixture-of-experts model with 8B parameters active per token for input and 16B for output.