DeepSeek introduces V4.1-Flash with native multimodal support

DeepSeek has released DeepSeek-V4.1-Flash, a new model built around a revised architecture and native visual understanding. The company published the announcement on September 10, 2026. It describes V4.1-Flash as the smallest model in its new architecture family, designed to improve text and agent performance while increasing inference speed and throughput. The release is a new product announcement, not simply a rebranding of the V4-Flash model already covered in DeepSeek’s earlier documentation.

DeepSeek says the model uses a 552-billion-parameter mixture-of-experts design. Its new Causal Encoder–Decoder architecture activates 8 billion parameters for input processing and 16 billion for output generation. The company also attributes the results to new pretraining methods and larger-scale reinforcement-learning post-training. DeepSeek’s page shows benchmark comparisons and says V4.1-Flash performs ahead of flagship models, including DeepSeek-V4-Pro. Those are provider-presented results; the page does not provide an independent evaluation, so they should be treated as claims to test rather than settled rankings.

The infrastructure changes are central to the release. DeepSeek says the model’s key-value cache needs one quarter of the high-bandwidth memory and one eighth of the SSD storage required by the previous generation. That matters because cache storage can become a significant cost in long-running agent workloads. Lower memory and storage requirements could allow providers to serve more concurrent requests or run longer contexts with less hardware, although the real effect will depend on prompts, batching, hardware and software implementations.

V4.1-Flash is now available through the DeepSeek API with native multimodal support. DeepSeek says the existing `deepseek-flash` model name can be used, while older V4-Flash and V4-Flash-Vision-Exp routes temporarily point to the new model. The company is phasing out V4-Pro and says Pro requests will route to V4.1-Flash at its rates from September 14. New pricing began at 04:00 UTC on September 10, with peak and off-peak rates continuing and off-peak usage priced at half the peak rate.

For AI users and makers, the release combines capability, efficiency and compatibility changes in one step. Native image understanding broadens possible agent workflows, while smaller caches could improve the economics of high-volume use. Teams should still measure visual accuracy, latency, cost and tool-use reliability on their own tasks, and should note the operational change when older model names are routed to a newer system.