DeepSeek has officially released the DeepSeek V4.1 Flash model, the smallest entry in its new architecture series featuring native multimodal visual understanding. Built on a Causal-Encoder-Decoder design, it is a 552B-parameter Mixture of Experts (MoE) model that activates only 8B parameters on the input side and 16B on the output side.

  • The asymmetric architecture significantly reduces inference costs compared to known models of the same size.
  • KV cache compression reduces HBM requirements to 1/4 and SSD requirements to 1/8 of the previous generation, lowering costs for agent-style tasks.
  • Benchmark testing shows V4.1 Flash surpasses the intelligence level of flagship models including DeepSeek V4 Pro.
  • The model is now live on the DeepSeek API under the name `deepseek-flash`, with older V4 Flash variants routed to it temporarily.
  • Pricing has been adjusted with peak/off-peak rates, and requests to `deepseek-v4-pro` will be routed to V4.1 Flash after September 14, 2026.

The new architecture aims to provide a higher capability ceiling, faster inference, greater throughput, and scalability to larger-parameter models while serving more users at lower cost.