What is DeepSeek V4.1 Flash?
DeepSeek V4.1 Flash is the smallest model in DeepSeek's new architecture family and its current Flash model. It replaced V4 Flash and V4 Flash Vision Exp, which DeepSeek has retired. It is a 552B-parameter mixture-of-experts model that reads text and images, writes text, and holds 1M tokens of context.
Only 8B parameters work on each token while it reads and 16B while it writes, a split DeepSeek built for agents that read far more than they write. You call it as deepseek-flash on DeepSeek's API, or download the MIT-licensed weights and run them yourself.
What changed from DeepSeek V4 Flash?
Almost everything: it is a new, larger model with image input, a much smaller memory cache and lower API prices, and V4 Flash itself is retired.
- Size
- 552B parameters in total against 284B, but 8B active while reading input instead of 13B.
- Images
- It reads images natively; V4 Flash was text only, with vision limited to an experimental version.
- Memory
- Its context cache needs about a quarter of the GPU memory and an eighth of the disk space V4 Flash used.
- Scores
- In DeepSeek's tests it rose from 82.7% to 90.6% on Terminal-Bench 2.1 and from 54.4% to 74.2% on DeepSWE.
- Old model names
- Calls to deepseek-v4-flash now run on V4.1 Flash and are billed at its price.
Who is DeepSeek V4.1 Flash best for?
DeepSeek V4.1 Flash is best for input-heavy agent and coding jobs, very long documents and codebases, teams that host their own models, and batch work run outside DeepSeek's peak hours.
Input-heavy agent work
It activates only 8B parameters per token while reading input, a design DeepSeek aimed at agents that read far more than they write.
Very long documents and codebases
The context window reaches 1M tokens, with up to 384K tokens of output through the API.
Teams that host their own models
MIT-licensed weights can be downloaded, changed and run on your own servers, which keeps prompts off DeepSeek's API.
Batch jobs outside peak hours
API prices halve outside weekday peak hours, so scheduled jobs pay $0.15 per million input tokens.
What's included with DeepSeek V4.1 Flash?
Adjustable thinking, image input, OpenAI and Anthropic API formats, very cheap cached input, tool calls with JSON output, and 2,500 concurrent requests per account.
- Adjustable thinking
- Thinking is on by default at high effort; you can set low, high or max, or switch it off for quicker replies.
- Image input
- It reads JPEG, PNG, GIF and WebP images sent inline, by URL or through the Files API, at up to 1,024 tokens per image.
- OpenAI and Anthropic formats
- The API accepts Chat Completions, Responses API and Anthropic-style requests, so tools such as Claude Code and OpenCode can use it directly.
- Very cheap cached input
- Cache hits cost $0.006 per million tokens at peak and $0.003 off-peak, 2% of the normal input price.
- Tool calls and JSON output
- Tool calling and JSON mode are standard, with chat prefix completion and fill-in-the-middle completion in beta.
- High concurrency
- Each account can run 2,500 requests at once, five times the V4 Pro limit, and DeepSeek raises it on request at no charge.
How much does DeepSeek V4.1 Flash cost?
DeepSeek V4.1 Flash is free to download under the MIT license, and DeepSeek's API charges $0.30 per million input tokens and $1.20 per million output tokens at peak hours, half that off-peak.
- Download from Hugging Face
- 552B parameters, 8B active while reading input and 16B while writing
- 1M-token context, image and text input
- Cached input $0.003 per million tokens
- All hours outside peak, including weekends
- Model name deepseek-flash
- Cached input $0.006 per million tokens
- Weekdays 01:00 to 04:00 and 06:00 to 10:00 UTC
- Up to 384K output tokens
Off-peak API rates are half the peak rates, and Chinese public holidays count as off-peak. The API accepts images, supports thinking and non-thinking modes, and allows 2,500 concurrent requests. Requests to the older deepseek-v4-flash model names now run on V4.1 Flash at the same price. Charges come out of a prepaid balance. Prices are in US dollars. Full pricing on api-docs.deepseek.com. We checked these prices on .
How do readers rate DeepSeek V4.1 Flash?
No one has rated DeepSeek V4.1 Flash yet. Be the first.
Sign in to rate. One rating per account.
DeepSeek V4.1 Flash pros and cons
It is fast, cheap and good with images, but users say it hallucinates more than its rivals, acts too eagerly, and has grown too big for most home setups.
Pros
- Fast: users on Hacker News and Reddit keep coming back to its speed, and several moved to it from GLM-5.3-Flash for that reason.
- Cheap to run for long agent sessions, since cache hits cost $0.003 per million tokens off-peak and users report cache hit rates near 99%.
- Image input works well for screenshots and visual checks, users say, which earlier Flash models could not do outside an experimental version.
- People doing security reviews report good results and few refusals, and DeepSeek's own table puts it first on CyberGym at 88.1%.
Cons
- Several users say it hallucinates more than GLM-5.3-Flash and needs a stronger model or a tight plan to check its work; DeepSeek itself says a gap remains on the hardest tasks.
- It is very eager: users report it acting before being asked and, in one case, running risky system commands, so it belongs in a sandbox when it runs unattended.
- At 552B parameters it is almost twice the size of V4 Flash (284B), and users who ran the old model at home say the new one is too big for their hardware.
- On DeepSeek's own API, data is stored in China and its privacy policy allows training on it unless you opt out.
Is DeepSeek V4.1 Flash open source?
Yes. DeepSeek V4.1 Flash's code is public under the MIT licence, so you can run it on your own server.
What are the best DeepSeek V4.1 Flash alternatives?
DeepSeek V4.1 Flash is an open source alternative to Jev, MiniMax M3, Gemini 3.5 Flash-Lite and 14 more.
JevPaidTypeSafe's System One model for fast, structured decisions in software automation.
MiniMax M3FreeMiniMax's multimodal coding and agent model with a 1M context and computer use.
Gemini 3.5 Flash-LiteFreeGoogle's low cost Gemini 3.5 model for high volume tasks and subagents.
GLM-5.3FreeChoose it when the hardest agent tasks matter more than speed; DeepSeek's own table puts it ahead on Terminal-Bench 4.0 (37.9% against 31.2%), at $1.40 / $4.40 per million tokens.
GPT-5.4 mini & nanoPaidOpenAI's small GPT-5.4 models with a 400K context and computer use.
Kimi K2FreeMoonshot's downloadable 1T-parameter MoE for coding and agents.
Questions about DeepSeek V4.1 Flash
What is DeepSeek V4.1 Flash?
DeepSeek V4.1 Flash is DeepSeek's open-weight model for coding and agent work: it reads text and images across 1M tokens, costs $0.30 per million input tokens at peak, and downloads free under MIT.
Who is DeepSeek V4.1 Flash for?
DeepSeek V4.1 Flash is best for input-heavy agent and coding jobs, very long documents and codebases, teams that host their own models, and batch work run outside DeepSeek's peak hours.
Is DeepSeek V4.1 Flash free?
Yes. DeepSeek V4.1 Flash has a free Open weights plan: download from Hugging Face.
How does DeepSeek V4.1 Flash compare with DeepSeek V4 Pro on price?
At peak hours DeepSeek V4.1 Flash costs $0.30 per million input tokens and $1.20 per million output tokens, against $1.32 and $3.96 for DeepSeek V4 Pro. Flash also accepts images and allows 2,500 concurrent requests, while V4 Pro does not take images and is limited to 500.
Is DeepSeek V4.1 Flash open source?
Yes. DeepSeek V4.1 Flash's code is public under the MIT licence, so you can run it on your own server.
What are the best DeepSeek V4.1 Flash alternatives?
The closest alternatives on Recaply are Jev, MiniMax M3, Gemini 3.5 Flash-Lite, GLM-5.3, GPT-5.4 mini & nano and Kimi K2.
Where can I try DeepSeek V4.1 Flash?
On its official website, huggingface.co.
When are DeepSeek's off-peak hours?
DeepSeek's peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays. Every other hour, including weekends and those holidays, is off-peak, when DeepSeek V4.1 Flash API prices drop by half.
Does deepseek-v4-flash still work on the API?
Yes, for now. DeepSeek has retired V4 Flash and V4 Flash Vision Exp, but calls to deepseek-v4-flash and deepseek-v4-flash-vision-exp are served by V4.1 Flash at its price. DeepSeek asks developers to switch to the model name deepseek-flash.
Does DeepSeek train on what I send to its API?
It can. DeepSeek's privacy policy says it stores data in China and uses it to train and improve its models, and it lists a right to opt out of training. Running the MIT-licensed weights on your own servers keeps prompts away from DeepSeek entirely.
Can I use DeepSeek V4.1 Flash for commercial work?
Yes. The weights are under the MIT License, which allows commercial use, and DeepSeek's API terms give you the rights to its outputs, including for your own products or for training other models.
Our verdict
DeepSeek V4.1 Flash, DeepSeek's current Flash model, is one of the cheapest ways to run a capable coding or agent model, as long as something checks its work.
It beats V4 Pro on DeepSeek's agent tests at under a third of the price, reads images, and is fast; users back that up but report more hallucinations than GLM-5.3-Flash.
- Skip it for
- Unattended agent runs with no review, and data that must stay out of China unless you host the weights yourself.
- Check first
- On DeepSeek's API, schedule big jobs for off-peak hours, when every price halves.
DeepSeek-V3.2
DeepSeek V4 Pro
DeepSeek-V3.2-Speciale
Qwen3.5 Small
MiMo-V2.6-Flash
Step 3.7 Flash