DeepSeek released DeepSeek-V4.1-Flash on Sept. 10, 2026. The company published the model's trained files on Hugging Face, so anyone can download it and, with powerful enough computers, run it themselves. The files come under the MIT license, a license that lets anyone use, change and share them free of charge, and the download is large, about 510 GB. For developers who pay to use DeepSeek's models online, the launch also came with a plan to move V4-Pro users onto Flash, which the company has since backed away from.
V4.1-Flash reads images as well as text and answers in text, and it can take in about 750,000 words at once. DeepSeek says the core of the model has 552 billion parameters, the numbers a model tunes as it learns, and that its design switches on only about 8 billion of them for each piece of text it reads, and 16 billion for each piece it writes. Switching on only part of a model at a time lets it run faster than its full size suggests, though it still needs enough memory to hold the whole model.
DeepSeek's technical report focuses on shrinking the memory a model keeps while it reads long documents. The company says the new design needs about one-quarter as much of that memory for each piece of text as its V4-Flash model did.
Developers can also use the model through DeepSeek's API, a connection that lets developers plug a service into their own apps, where it is called deepseek-flash. DeepSeek charges by the amount of text and cuts its prices outside peak hours. At peak times, sending the model about 750,000 words of new text costs $0.30, and getting that much text back costs $1.20. Off-peak, those prices fall to $0.15 and $0.60. Repeated text that the service has stored from an earlier request costs far less: $0.003 off-peak and $0.006 at peak.
In its launch announcement on Sept. 10, 2026, DeepSeek said that from 04:00 UTC on Sept. 14, all requests for deepseek-v4-pro would go to V4.1-Flash and be billed at Flash rates. Its pricing page and changelog now say DeepSeek will keep offering V4-Pro after Sept. 14 with billing unchanged. The launch page still describes the Sept. 10 plan.
DeepSeek describes the new model as ahead of V4-Pro, but DeepSeek's own results table still shows V4-Pro scoring higher on some exams. Its launch text also does not say whether the model is available in DeepSeek's chat app or on its website.