DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17th with a one-sentence description: “a minor update to DeepSeek-V4.1-Flash.” That framing undersells what the benchmark table shows.
The original V4.1 Flash landed in September with strong numbers for a model at its price point. The 1017 revision keeps the same architecture — 763B total parameters, roughly 8B active per token, 1M context, MIT license — but the numbers on the agentic suite have moved substantially:
| Benchmark | Flash-1017 | Flash |
|---|---|---|
| Terminal-Bench 4.0 | 42.3 | 31.2 |
| ProgramBench | 29.1 | 20.3 |
| Apex Bench v1.1 | 72.1 | 64.6 |
| HLE w/o tools | 42.6 | 36.8 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 |
| DeepSWE v1.1 | 77.0 | 74.2 |
Terminal-Bench is a good proxy for “can this model actually operate a terminal without losing track of state,” and going from 31.2 to 42.3 is not cosmetic. ProgramBench, which tests multi-file code reasoning at repository scale, went from 20.3 to 29.1 — a 43% jump. These are the kinds of improvements that change whether a model is usable as an autonomous agent or whether it requires babysitting.
The improvements on Apex Bench and Automation-Bench point in the same direction: the model is better at sustained multi-step execution. HLE (Humanity’s Last Exam, no-tools variant) going from 36.8 to 42.6 suggests the base reasoning also improved, not just the task-specific fine-tuning.
All benchmarks are vendor-reported. DeepSeek publishes detailed model cards but doesn’t share training details, so there’s no independent verification of what changed between Flash and Flash-1017.
What’s notable about this update is the distribution of gains. DeepSWE v1.1 improved by only 2.8 points, while Terminal-Bench improved by 11.1. That pattern suggests the revision targeted the long-horizon execution side of the agentic stack — the part that requires maintaining coherent state across many steps — rather than pure coding ability. That’s the harder problem, and it’s the one that matters more for practitioners who are running agents on real infrastructure rather than running evaluations.
DeepSeek’s pricing structure, combined with the MIT license on the weights, makes this model practically available at a level that frontier-tier models aren’t. The combination of 1M context, open weights, and these numbers puts it in a useful position for teams building agentic systems where cost per token matters and where running the model locally is a genuine option. The “minor update” label seems to reflect engineering humility more than benchmark reality.