DeepSeek V4.1-Flash: 4× Leaner KV Cache and Stronger Coding

DeepSeek has released V4.1-Flash — a multimodal model with 552 billion parameters, of which only 16 billion are active at any given moment. The KV cache — the working memory the model uses during generation — is four times smaller than in its predecessor. On the DeepSWE coding benchmark, the model outperformed Opus 5 and GPT-5.6 Sol and was released under the open MIT license.

The less memory a model needs at runtime, the cheaper it is to run hundreds of agents in parallel. V4.1-Flash is squarely aimed at agentic scenarios: open-source code plus a compact active layer make it a genuine competitor to expensive proprietary models.