Source: deepseek.com - An MoE model with 552 billion parameters and native vision.
- Only 8 billion parameters are activated at the input.
- KV cache is down to 890 bytes per token.
- API is cheaper by up to 60%, V4 Pro ends on September 14.
DeepSeek launched V4.1 Flash on September 10 and did something that you rarely see in the world of large language models: the smallest member of the new family beat its own flagship in tests. The Chinese company therefore immediately announced that V4 Pro will be shut down on September 14. Instead, users will get a cheaper and faster model that also, for the first time, natively understands images.
The smallest model in the family that outgrew the flagship
V4.1 Flash is a MoE model with 552 billion parameters built on a completely new Causal-Encoder-Decoder structure. In simplified terms, the model forms a 20-layer causal encoder followed by a 20-layer decoder, where the input and output sides are not symmetrical. When reading the input, only 8 billion parameters are activated; when generating the response, 16 billion. That’s exactly where the significantly lower price comes from compared to models of comparable size.
The second new feature is native multimodal vision. The model processes images directly, without the helper add-on DeepSeek previously offered as a separate experimental version. The context window goes up to one million tokens and the company released the weights on Hugging Face under the MIT license, so anyone with the right hardware can download and run the model.
Where V4.1 Flash wins and where it still falls short
According to DeepSeek’s official results, the new model surpasses V4 Pro especially in areas where it has to act and program on its own. On Codeforces it scored 3471 vs 3348, on the DeepSWE v1.1 test it got 74.2 points vs 62.7, and on Terminal-Bench 3.0 it jumped to 30.0 from the previous 11.8. The same gap shows in Automation-Bench and CyberGym tests that measure tool work and security tasks.
However, it’s not a win across the board — and it’s worth saying plainly. In the knowledge test GPQA Diamond, V4.1 Flash slightly lags behind the older V4 Pro (90.9 vs 92.4). And in the demanding Humanity’s Last Exam, the gap is even more noticeable: 36.8 vs 42.7. Versus the Western top, i.e., GPT 5.6-Sol and Claude Opus 5, it only leads in part of the disciplines. The strength of the new model is the price-to-performance ratio in agentic tasks, not an absolute record in knowledge.
KV cache shrank to 890 bytes per token
The most technically interesting part of the announcement is about memory. DeepSeek dramatically reduced the so-called KV cache — the place where the model keeps the conversation context so far. Compared to the previous generation, the need for fast memory HBM dropped to a quarter and the requirements for SSD storage fell to an eighth.
In absolute numbers, it’s 890 bytes per token. The company’s first model requested 389,120 bytes for the same token, so in under three years this item shrank 437x. It sounds abstract, but in agentic tasks, cache makes up a large part of the bill. The model keeps the entire work history in it, and you’re charged for every reread of it.
Prices go down, most of all at the input side from cache
The hardware savings are reflected directly in the pricing. Input tokens that are found in cache are 60% cheaper, input without cache is about one third cheaper, and output is roughly 11% lower. Off-peak, one million input tokens costs $0.15 and one million output tokens costs $0.60 — which is roughly €0.13 and €0.52. During peak hours, the rates are exactly doubled.
This is pretty good news for us. According to DeepSeek, the peak rates last on weekdays from 1:00 to 4:00 and from 6:00 to 10:00 UTC, which in our summer time translates to 3:00 to 6:00 and 8:00 to 12:00. So a Slovak developer who sits down at their computer in the afternoon or evening will automatically pay the reduced rate. The same applies all weekend and on holidays. New prices apply starting 10. September.
V4 Pro is being shut down on September 14
In the API, the new model is simply called deepseek-flash. The company has temporarily redirected the older names deepseek-v4-flash and deepseek-v4-flash-vision-exp to it so nothing breaks in existing apps. Tencent partners were the first to roll out the new feature, with WorkBuddy and CodeBuddy services, along with the OpenCode developer tool.
A bigger change is coming for users of the paid V4 Pro version. After September 14, all requests for this model will be routed to V4.1 Flash and charged according to its cheaper rates, until the company introduces V4.1 Pro. The Chinese site IT Home, meanwhile, pointed out that the original shutdown date was earlier, and DeepSeek moved it at the last moment. It was informed by an email the company sent out to users of the interface.
| - | DeepSeek V4.1 Flash |
|---|---|
| Model type | MoE / natively multimodal |
| Parameters | 552 billion |
| Active parameters | 8 billion in / 16 billion out |
| Architecture | Causal-Encoder-Decoder / 40 layers |
| Context | up to 1 million tokens |
| KV cache | 890 bytes per token |
| API name | deepseek-flash |
| License | MIT / weights on Hugging Face |
Follow the China Planet website, subscribe to our monthly newsletter, and be sure to check out our Facebook page, the China Planet group, and join our Telegram, where we add the best offers and coupons from China every single day.
Don’t miss the best coupons 📩
Subscribe and we’ll email you new coupons before they expire. We only send it once a month — no spam.
