DeepSeek V4.1 Flash launch

DeepSeek V4.1 Flash is a 552-billion-parameter Mixture-of-Experts model released on September 10, 2026, built on a Causal Encoder-Decoder architecture. Its asymmetric design activates only 8 billion parameters on the input side and 16 billion on the output side, which DeepSeek said results in 'significantly lower cost than other known models at a similar scale.' The model carries native visual understanding, a 1-million-token context window, and was released as an open-weight model on Hugging Face. DeepSeek characterized it as the smallest model in a new architecture family, '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. DeepSeek claimed benchmark results ahead of DeepSeek-V4-Pro, though one report noted the benchmarks were selectively chosen. 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. The model is specifically built to reduce cache burden in long agent workloads. Databricks made DeepSeek 4.1 Flash available on its platform the same day it shipped, routing all calls through its Unity Gateway for unified access, spend, and performance monitoring.