DeepSeek launches V4.1 Flash with lower costs and stronger coding performance

DeepSeek has introduced V4.1 Flash, an open multimodal model with support for context windows of up to one million tokens.

The model uses a mixture-of-experts architecture with 552 billion parameters, but only a fraction of them are active at any given time. DeepSeek says its new Causal Encoder-Decoder design uses just 8 billion active parameters when processing input and 16 billion during generation, reducing compute requirements for more demanding agentic workloads.

Memory use has also been cut substantially. Compared with the previous V4 Flash, V4.1 Flash uses four times less KV-cache memory and takes up eight times less storage space. DeepSeek says it achieved this by compressing the cache to 890 bytes per token and recomputing some states on the fly instead of continuously writing them to disk.

In benchmark testing, the new model outperformed the larger DeepSeek V4 Pro. It scored 74.2 on DeepSWE v1.1, 90.6 on Terminal Bench 2.1, and reached a Codeforces rating of 3471, with the strongest gains appearing in coding and agent-based tasks.

V4.1 Flash also adds a reasoning-effort slider from 1 to 100, giving users more control over the trade-off between response speed and accuracy. DeepSeek says the model was pretrained from scratch on its own data, while distillation was used only during the final post-training stage through OPD.

DeepSeek has also cut API pricing alongside the release. During peak hours, one million input tokens cost between $0.006 for cached input and $0.30 for standard input, while output is priced at $1.20 per million tokens. Off-peak rates are half that, at $0.003 for cached input, $0.15 for standard input and $0.60 for output.

V4.1 Flash is already available to developers under an MIT license, with model weights and encoding libraries published in the project repository. DeepSeek is also retiring the original V4 Flash and V4 Flash Vision Exp, and from September 14, 2026, requests sent to V4 Pro will be redirected to V4.1 Flash at the lower pricing tier.