DeepSeek-V4.1-Flash-1017: MIT-licensed open-weight update to V4.1-Flash

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, with open weights on Hugging Face and ModelScope. Its news page says the model has stronger benchmark results, faster task completion and shorter outputs than the previous version. It is live on DeepSeek's website, app and API. DeepSeek's API docs say the `deepseek-flash` model name is now served by 1017, and the Change Log says the API calling method and pricing are unchanged. The Hugging Face model card calls 1017 a minor update to DeepSeek-V4.1-Flash and claims gains in reasoning, coding, agentic tasks and visual benchmarks. The weights are MIT licensed and the repo is not gated. The 1017 card gives no parameter count. The card for the previous V4.1-Flash (Hub repo created September 10, 2026) lists 552B backbone parameters, 8B active per token in prefill and 16B in decode, 1 shared and 384 routed experts with 6 routed experts active, and Engram conditional memory of 196B parameters. Hugging Face's safetensors metadata for the 1017 repo counts about 763B parameters in total. DeepSeek does not state that figure, and it is a different measure from the 552B backbone figure. The 1017 config.json is byte-identical to the predecessor's. It sets 40 hidden layers, `max_position_embeddings` of 1,048,576, and FP8 weights with FP4 experts, and it includes a vision encoder. The card recommends a 1M-token context window. DeepSeek ships no GGUF. Its only inference guidance is a reference implementation run with torchrun, which DeepSeek says is not a production serving engine. Its examples use 8-way model parallelism. The vLLM, SGLang and Docker Model Runner options on the Hub page are generated by Hugging Face, not listed by DeepSeek. Community GGUF, EXL3, NVFP4, MXFP4 and MLX builds are already on the Hub.