Release dates
— DeepSeek V4.1 Flash reached general release and entered the API under the recommended name `deepseek-flash`; the old V4 Flash and Vision Exp models were retired and temporarily route to the new model. ↗
DeepSeek-V4.1 Flash
The original DeepSeek V4.1 Flash snapshot, released on 2026-09-10: a 552B MoE with native text-and-image input, text output, 1M context, and up to 384K output. DeepSeek’s web, app, and API now use the separately recorded 1017 version.
From 2026-10-17, DeepSeek’s web, app, and API use V4.1-Flash-1017; `deepseek-flash` and legacy compatible names are served by the new version. This record describes the original September snapshot, whose weights remain available in its original repository. DeepSeek’s pricing page updated on 2026-09-14 confirms V4 Pro remains a separate API and was not routed to Flash as previously planned.
MODEL NUMBER107
- First public date
- 2026-09-10
- Organization
- DeepSeek
- Family
- DeepSeek V / R
01
Specifications
Only published specifications are shown.
- Total parameters
- 552B↗
- Active parameters
- 16B↗
- Architecture
- MoE↗
- Parameter scope
- DeepSeek distinguishes a 552B backbone, 8B active per token during input prefill, and 16B during decode; activeParameters uses the decode-side figure and does not treat the combined tensor size shown by Hugging Face as the backbone.↗
- Context
- 1,000,000 tokens↗
- Maximum output
- 384,000 tokens↗
- Input modalities
- Text · Image↗
- Output modalities
- Text↗
- Open weights
- Yes↗
- License
- MIT↗
- Status
- Current↗
02
Overview
Why it mattered
The original open-weight V4.1 Flash checkpoint remains available; DeepSeek later moved the stable API model ID to the separately released 1017 checkpoint.
PreviousDeepSeek-V4-Flash-0731284B · 1M ctx
CurrentDeepSeek-V4.1 Flash552B · 1M ctx
04
Architecture & capabilities
Attention
Causal Encoder-DecoderCompressed Sparse Attention 2
Model capabilities
VisionAgentCodingLong context
API features
Adjustable reasoningResponses API
05
Release history
[Unofficial beta] The temporary API endpoint `deepseek-v4.1-flash-expires-on-0910` is callable; included after user testing, not a formal public release.
View sourceDeepSeek V4.1 Flash reached general release and entered the API under the recommended name `deepseek-flash`; the old V4 Flash and Vision Exp models were retired and temporarily route to the new model.
View source06
Related models
Select a card to view the related model.
PredecessorDeepSeek-V4-Flash-0731View details →ReplacesDeepSeek-V4-Flash-0731The compatibility model name is temporarily routed to V4.1 FlashView details →ReplacesDeepSeek-V4-Flash-Vision-ExpThe compatibility model name is temporarily routed to V4.1 FlashView details →SuccessorDeepSeek-V4.1-Flash-1017View details →
07
Sources
01
DeepSeek V4.1 Flash 中间版本内测报道IT之家 · Published 2026-09-08 · Checked 2026-09-08
Reporting · Unconfirmed02DeepSeek V4.1 Flash：更强、更快、更普惠DeepSeek · Published 2026-09-10 · Checked 2026-09-10
Official · Verified03DeepSeek-V4.1-Flash model cardDeepSeek · Checked 2026-09-10
Official · Verified04DeepSeek models and pricingDeepSeek · Published 2026-08-21 · Checked 2026-09-14
Official · Verified05DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made RealDeepSeek · Published 2026-10-17 · Checked 2026-10-17
Official · Verified06DeepSeek API changelog and model guide: V4.1-Flash-1017DeepSeek · Checked 2026-10-17
Official · VerifiedSuggest a correction
Copy this template, add the field, proposed value, and primary source, and share it with the archive maintainer.