DeepSeek V4.1 Flash Fast

DeepSeek announced V4.1-Flash on September 9, 2026 as the smallest model in a redesigned architecture family, featuring a 552B-parameter Mixture-of-Experts backbone with a new Causal Encoder-Decoder design. It activates just 8B parameters for input and 16B for output, enabling faster inference and higher throughput than the prior generation. The model ships with native visual understanding for multimodal workloads. V4.1-Flash reaches a smaller KV cache footprint than the previous generation, using one-quarter the HBM and one-eighth the SSD storage, which materially reduces agent cache-hit costs. Tests by multiple parties place V4.1-Flash ahead of the V4-Pro flagship on performance, cost, speed, and total runtime. The weights and technical report are published as deepseek-ai/DeepSeek-V4.1-Flash on Hugging Face for community inference support.

DeepSeek's V4.1-Flash release pairs the 552B-parameter MoE backbone with a one-million-token context window and native multimodal vision, positioning the model for high-frequency batch generation and cost-sensitive agent loops. Early community measurements cited around launch pointed to inference speeds near 400 tokens per second at Flash-tier pricing. Vendor-reported instruct benchmarks released with V4.1-Flash include Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2, HumanEval pass@1 at 79.4, and strong multimodal scores on DocVQA and CVBench, running with maximum reasoning effort and DeepSeek's own harnesses. Independent evaluators had not fully replicated these numbers at launch, so the figures should be treated as the vendor's release claim rather than third-party confirmation.