DeepSeek Releases V4.1 Flash, a 552B MoE With Native Vision

DeepSeek released V4.1 Flash on September 10, 2026, describing it as the smallest model in a new architecture family with native visual understanding. The model is a 552-billion-parameter Mixture-of-Experts system built on a Causal-Encoder-Decoder architecture, with an asymmetric design that activates only 8 billion parameters on input and 16 billion on output. DeepSeek said it is designed for greater capability, faster inference, higher throughput, and scaling to larger models. The model incorporates a new pre-training method and larger-scale reinforcement learning post-training, and is available on Hugging Face. Analyst Chris Porter said the model's KV cache is 437 times smaller than competing models and that this helps it beat GPT-5.6 Sol on several agentic benchmarks.