DeepSeek has published V4.1-Flash-1017 on Hugging Face, a minor update to V4.1-Flash. The repository and the model weights are released under the MIT licence. The model page lists 763 billion parameters, with tensors in BF16, F32, F8_E4M3 and I8 formats. It gives no release date; the AI Weekly news digest lists the model under 17 October.
The page says the update delivers "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks" and "reduces task execution time and output length, enabling users to complete tasks faster and at lower cost".
The scores below are those reported on the model page, compared with the previous V4.1-Flash. On the HLE benchmark without tools the new model scores 42.6 against 36.8. On Terminal-Bench 4.0 it scores 42.3 against 31.2, a test run with the Claude Code harness. DeepSWE v1.1, run with the Mini-SWE-V2 harness, goes from 74.2 to 77.0, and ProgramBench from 20.3 to 29.1. Apex Bench v1.1 rises from 64.6 to 72.1. With tools, BabyVision moves from 89.6 to 92.5 and Chartography from 78.9 to 83.0. The page also carries comparison columns for other models.
DeepSeek recommends a temperature of 1.0 and a top_p of 1.0, a 1-million-token context window, and a max_tokens setting of at least 256,000.
The release announces no new model API and ships no Jinja chat template. Prompts are encoded with a Python reference implementation, encoding.py, held in the repository. A separate tool, deepseek-recipe, consists of Rust libraries with Python bindings. It converts Messages, Chat Completions and Responses API requests into DeepSeek V4 and V4.1 prompts and parses the model's outputs. Running inference, executing tools and handling HTTP transport remain the caller's job.
The benchmark figures come from the model page itself.
