It's too fast.
On Thursday morning, DeepSeek V4.1 Flash officially went live on the web, apps and APIs.
According to official disclosures, V4.1 Flash has outperformed V4 Pro in key performance, response speed, and efficiency indicators, and adopts a new architecture and natively supports multi-models.
Specifically, V4.1 Flash is a MoE model with a total parameter size of 552B. It uses a completely new Causal-Encoder-Decoder architecture and uses a non-symmetrical computational design for input and output: the input side only activates 8B parameters and the output side activates 16B parameters, thus significantly reducing the actual reasoning costs while maintaining large model capacity.
In addition to the architectural adjustments, V4.1 Flash introduced new pre-training methods and performed larger reinforcement post-training. Eventually, it achieved performance improvements in multiple benchmark tests, with overall intelligence levels higher than many of the current flagship models, and a better balance between performance, efficiency and cost.
The model is open-source:
From the official benchmark results shown below, V4.1 Flash performed strongly in tests such as software engineering, cybersecurity environments and automation tasks.
On DeepSWE v1.1, CyberGym and Automation-Bench, it scored the highest in the chart, with DeepSWE only 0.2 points slightly ahead of Opus 5; but on Terminal-Bench 3.0, V4.1 Flash scored 30.0 points significantly below Opus 5's 43.3 points and GPT-5.6 Sol's 34.4 points.
For a more comprehensive benchmark test result, you can see that in addition to significant improvements in coding, software engineering and Agent tasks, V4.1 Flash still has a gap with top closed source models such as GPT-5.6 Sol or Claude Opus 5 in projects such as GPQA Diamond, Terminal-Bench 3.0/4.0, and ProgramBench.
DeepSeek was so confident in the 4.1 Flash that it announced early that it would cut the V4 Pro.
V4.1 Flash The other main line of this update is to further reduce the cost of reasoning.
The token structure of the Agent-class load is extremely slanted, often in the context of tens of thousands of tokens, tool returns, code warehouse forecasts, and tens to hundreds of token outputs. V4.1's input/outputs greatly optimize the total cost of improvements to this type of task.
According to official data, the KV Cache of the new generation of models shrank significantly, reducing the demand for HBMs to 1/4, compared to the previous generation, to 1/8 of the demand for SSDs. The KV Cache of the V4 generation is now only about 7% of V3.2.
It is worth noting that a model API would be concerned with SSD requirements, indicating that its caching layer would have used SSD as a tiered storage. This is the consistent route for the DeepSeek service layer.
Now that the KV readings are so small that they no longer dominate the display bandwidth, weight reading is the bottleneck, which is why the community detects 400 tokens/s, "speed does not decrease with context length".
After noon on September 14, V4 Pro's API will be offline, and all API requests directed to V4 Pro will be automatically routed to V4.1 Flash and processed at a flat rate following the Flash series price reduction: 0.02 USD for caching time input, 1 USD for caching time input, 4 USD for output; the peak price is twice the price for caching time input.
At the same time, large-scale C-end users have also received a major upgrade: on the web and App edges, the original fast, expert, mapping three modes have all been merged, accessed by V4.1 Flash, and can now start deep thinking + search with the most advanced big models.
In addition, some third-party Agent tools and AI platforms have simultaneously launched new models, such as Workbuddy.
Previously, Op.