Although it appears to be a minor update, the October revision of DeepSeek’s Flash model gains nearly 6 points on HLE and 11 on Terminal-Bench, whilst producing shorter responses.
DeepSeek filed on Hugging Face DeepSeek-V4.1-Flash-1017, a revision of its Flash model with mixture-of-expert architecture. According to the model sheet, it has 763 billion parameters in total, accepts up to a million context tokens and is distributed under MIT license, so commercially reusable without any particular restrictions. The Chinese laboratory presents it as a minor update, which nevertheless improves reasoning, code, agentic tasks and vision while reducing execution time and output length.
Net earnings in agentics
The differences published with the previous version are significant: 42.6 versus 36.8 on Humanity's Last Exam, 42.3 versus 31.2 on Terminal-Bench 4.0, 77.0 versus 74.2 on DeepSWE v1.1 and 29.1 versus 20.3 on ProgramBench. Weights can be served with Transformers, vLLM or SGLang, and quantified versions already exist for llama.cpp, Ollama and LM Studio.
Read the full article on the original source
Go to the source →