Details
- DeepSeek announces V4.1-Flash as the smallest model in its new architecture family, featuring native visual (multimodal) understanding.
- The model uses an asymmetric 552B-parameter Mixture-of-Experts design with a causal encoder–decoder, activating only 8B parameters for input and 16B for output to cut compute costs.
- Compared with the previous V4 generation, V4.1-Flash reduces KV cache requirements to one-quarter of the HBM and one-eighth of the SSD storage, significantly lowering cache-hit driven agent costs.
- V4.1-Flash is now live on the DeepSeek API under the model name deepseek-flash, with native multimodal support; legacy models V4-Flash and V4-Flash-Vision-Exp are retired, with their endpoints temporarily routed to V4.1-Flash for compatibility.
- DeepSeek keeps peak and off-peak API pricing, with off-peak rates set at 50% of peak, encouraging users to schedule flexible workloads during cheaper off-peak windows.
- The company plans to collaborate closely with the open-source community to add V4.1-Flash inference support and invites teams planning large-scale deployments on 2,000+ GPUs with storage clusters to engage on deployment options.
- DeepSeek reports that new pre-training methods combined with larger-scale RL post-training deliver benchmark results ahead of its prior generation, aiming to offer greater capability, faster inference, and higher throughput while scaling to larger models.
Impact
By pairing a very large MoE backbone with a small number of active parameters, DeepSeek positions V4.1-Flash as a cost-efficient workhorse model that can compete with other mid-tier frontier APIs on price-sensitive workloads. The sharply reduced KV cache footprint directly targets growing agent and memory costs, while native multimodal support and open-source collaboration could broaden adoption among developers looking for scalable, GPU-efficient deployments.
