AI news · October 18, 2026
DeepSeek V4.1 Flash tightens the open model for agent work
DeepSeek V4.1 Flash 1017 reports 77.0 on DeepSWE v1.1 and a 1M-token context; run a small agent trace to see if shorter output lowers cost.
- reported parameters
- 763B
- context tokens
- 1M
- DeepSWE v1.1 score
- 77.0
DeepSeek's Hugging Face release for V4.1-Flash-1017 describes a minor update to V4.1-Flash that improves reasoning, coding, agent work, and visual tests while reducing task time and output length. The card lists 763B parameters, a 1M-token context window, MIT weights, and local paths for Transformers, vLLM, SGLang, and Docker.
The vendor table reports 42.3 on Terminal-Bench 4.0, 77.0 on DeepSWE v1.1, and 83.0 on Chartography with tools. These are DeepSeek-run comparisons, so they show where to start, not a neutral score. The practical angle is the combination of a large context and fewer output tokens: measure time and bill on your own agent trace before changing a default.
What you can do with it
Run five representative tasks with the same prompt and tools on your current model and V4.1-Flash-1017. Record accepted result, output tokens, retries, tool errors, wall time, and cost. Keep the model only if it improves accepted work after review.
Our take
The update is more useful than a headline score because it targets agent time and output length. A 1M context is not a reason to send every document; long prompts still cost money and can hide stale context. Start with a short trace and compare the full task, not just the first answer.
Links Try itDeepSeek API docs
Source: Hugging Face / DeepSeek — Made With Models writes the brief; the reporting is theirs.