DeepSeek V4.1 Flash (Deep Infra)

DeepSeek V4.1 Flash is positioned as a compact, efficiency-focused member of a new DeepSeek architecture family, and is described as the first Flash-tier variant to ship with native vision and multimodal understanding. Its design uses a Mixture of Experts setup with 552B total parameters, activating only 8B parameters for input and 16B for output, built on what the source calls a Causal Encoder Decoder architecture. Open weights are published on Hugging Face, making the model usable outside of hosted APIs and attractive for teams that want to self-host or fine-tune.

Practical strengths emphasized for the model include frontier-leaning intelligence at modest active parameter counts, a long context window of 1M tokens, and a maximum output of the cataloged API limit tokens, with both thinking and non-thinking modes plus tool calling and structured JSON output. The source claims it surpasses DeepSeek V4 Pro on performance, cost, speed, and task completion time, and reports a score of 88.1 on CyberGym, while a heavily compressed KV cache of roughly 890 bytes per token is highlighted as a way to lower cache hit costs in agent workloads. It is served in FP8 on Deep Infra with prompt caching, making it well suited to coding agents, long-context retrieval, high-throughput batch processing, and multimodal assistants.