The Hugging Face model card is timestamped 2026-10-17T05:52:18Z, 13:52 in Beijing on 17 October. That is before this morning edition’s window, 16:00 in Beijing on 17 October through 08:00 on 18 October, and still inside the past 24 hours. This draft uses the card’s own clock, not a later rewrite. DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash. The card says it brings substantial improvements on reasoning, coding, agentic tasks, and visual benchmarks, and that it reduces task execution time and output length so users can finish faster and at lower cost. It does not say how much time, how much output, or what the price is. With no multiplier, none is added.

[1]
A closed card, charcoal cover, muted blue edge
A closed card lying flat, the cover painted charcoal, the side edge muted blue., AI-generated illustration, not a news photograph

The six columns are this release, the previous Flash, K3, GLM-5.3, GPT-5.6 Sol, and Opus 5. The numbers are the card’s own table, not an independent retest. On HLE without tools, this release scores 42.6. The 44.0 in parentheses carries a dagger. The card says the dagger marks the text-only subset. The previous Flash is 36.8 (39.1). K3 is 43.5. GLM-5.3 is 42.0, also marked as the text-only subset. GPT-5.6 Sol is 44.5. Opus 5 is 56.3. This release is above the previous Flash, below Sol, and below Opus 5. The parenthetical subset is not the same exam as an unmarked 43.5, so 44.0 is not lined up against that 43.5 as one score.

[1]

Terminal-Bench 4.0 uses the Claude Code harness. This release scores 42.3, against 31.2, 12.6, 37.9, 39.9, and 51.8. It is above every column except Opus 5. DeepSWE v1.1 uses the Mini-SWE-V2 harness. This release scores 77.0, against 74.2, 67.5, 66.9, 73.0, and 74.0. That column is above the other five, including Opus 5. ProgramBench is 29.1, against 20.3, 17.5, 19.0, 23.0, and 37.0. Above the previous Flash and Sol, below Opus 5. Apex Bench v1.1 is 72.1, against 64.6, 47.8, 56.6, 54.3, and 65.8. That column is above the rest of the table. The card says Terminal-Bench uses Claude Code and DeepSWE uses Mini-SWE-V2. On another harness, these scores do not travel unchanged.

[1]

The repository and the weights are under the MIT license. The listed model size is 763B parameters. The recommended sampling settings are temperature 1.0, top_p 1.0, a context window of 1M tokens, and max_tokens of at least 256K. Those are recommended settings, not benchmark scores. This release does not include a Jinja-format chat template. The encoding implementation and the local inference steps are not copied into this draft. The table also lists NL2Repo, Automation, Agents’ Last Exam, and several visual scores. They do not change the judgment above: several agent columns are above the previous Flash and above some larger models in the table, while HLE, Terminal-Bench, and ProgramBench remain below Opus 5.

[1]