DeepSeek V4.1 Flash: 552B open MoE under MIT with KV cache cut to a quarter
In one sentence On September 10 DeepSeek released V4.1 Flash, a 552B multimodal MoE with 8B active parameters for input and 16B for output, a 1-million-token context and an MIT license. Versus the previous generation it claims a quarter of the HBM and an eighth of the SSD storage for the KV cache.
DeepSeek released V4.1 Flash, a new AI model with downloadable weights under the MIT license, so it can be used freely, commercially too. It understands text and images and can handle very long documents, up to one million tokens.
It is a very large model, but uses only a small slice of itself for each word: that is the mixture-of-experts (MoE) trick, which makes such models cheaper to run. The main news is the working memory the model keeps during long conversations: DeepSeek says it has cut it to a quarter of the previous generation.
For API users, DeepSeek also confirmed half-price rates during off-peak hours. Claimed performance beats the company's bigger V4-Pro model, but these are the vendor's numbers.
Companies
DeepSeek
Tools
DeepSeek-V4.1-Flash, vLLM, SGLang
Tags
Sources