ڈیپ سیک V4.1 Flash

A multimodal Flash successor for reasoning, coding and agent workflows, with native image input and a million-token context.

V4.1 Flash introduces DeepSeek's Causal Encoder-Decoder architecture and native visual understanding. It is intended for reasoning and agent workflows where throughput and the cost of repeatedly reading context matter alongside answer quality.

The design assigns different amounts of computation to reading and writing: eight billion active parameters for input and sixteen billion for output. DeepSeek attributes its capability gains to new pretraining methods and larger-scale reinforcement learning, and reports results ahead of V4 Pro on its evaluations.

The release also reduces the memory and storage required by the KV cache. DeepSeek reports one quarter of the previous generation's HBM footprint and one eighth of its SSD requirement. That efficiency is particularly relevant to agents that revisit long contexts across many steps; actual serving limits and billing depend on the selected host.