DeepSeek Launches V4.1-Flash: 552B MoE with Visual Support
Original: DeepSeek v4.1 Flash
Why This Matters
A 552B open-weight MoE model with reduced KV cache costs signals a new efficiency benchmark for large-scale AI inference.
DeepSeek announced DeepSeek-V4.1-Flash on September 10, 2026, a 552B-parameter MoE model featuring a new Causal Encoder–Decoder architecture with only 8B active parameters for input and 16B for output, plus native multimodal (visual) understanding.
DeepSeek unveiled DeepSeek-V4.1-Flash, the smallest model in its new architecture family, designed for faster inference, higher throughput, and scalability to larger models. The model uses a 552B-parameter Mixture-of-Experts (MoE) structure with a novel Causal Encoder–Decoder architecture that activates just 8B parameters for input and 16B for output, reducing compute costs significantly. KV cache requirements are dramatically reduced versus the prior generation: HBM usage drops to 1/4 and SSD storage to 1/8, cutting agent infrastructure costs. The model is immediately available via the DeepSeek API under the model ID 'deepseek-flash', with native multimodal support. Previous models V4-Flash and V4-Flash-Vision-Exp are retired, though their API endpoints temporarily route to V4.1-Flash for compatibility. DeepSeek also announced lower API pricing, with off-peak rates set at 50% of peak rates. The model weights have been released openly on Hugging Face at deepseek-ai/DeepSeek-V4.1-Flash, and DeepSeek stated it will collaborate with the open-source community on inference support and explore additional deployment options.