DeepSeek V4.1 Flash: official release links and benchmark caveats
Collects the launch thread, release notes, MIT-licensed HF weights, ModelScope mirror and Search & Fetch docs, noting all scores are DeepSeek's own and Opus 5 still leads on HLE, Terminal-Bench and others.
Proof: Launch thread (Oct 17): https://x.com/deepseek_ai/status/2111375640717033777 Release notes: https://api-docs.deepseek.com/news/news261017 Weights, benchmark table, MIT license: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017 ModelScope mirror: https://modelscope.cn/models/deepseek-ai/DeepSeek-V4.1-Flash-1017 Search & Fetch docs: https://api-docs.deepseek.com/guides/search Every score here is DeepSeek's own. Opus 5 still leads on HLE, Terminal-Bench, ProgramBench, NL2Repo, Chartography, BabyVision and ZeroBench.
This post quotes another post: x.com/deepseek_ai/status/2111375640717033777