Model Details
DeepSeek-V4.1-Flash-1017
Quick answer: DeepSeek-V4.1-Flash-1017 is a minor, open-weight (MIT-licensed) update to DeepSeek-V4.1-Flash, delivering gains across reasoning, coding, agentic, and visual benchmarks while reducing task execution time and output length — advancing the performance–cost Pareto frontier for agent harnesses. It carries forward the same 1M-token context and Causal Encoder-Decoder (CED) MoE architecture.
At a Glance
Where DeepSeek-V4.1-Flash-1017 leads
- DeepSWE v1.1: 77.0% resolved (vs 74.2% for the base V4.1-Flash)
- NL2Repo-Bench: 68.2% (vs 65.4% base)
- Agents' Last Exam: 35.4% (vs 31.8% base)
- Terminal-Bench 4.0: 42.3% — a large jump over the base model's 31.2%
Where it lags
- Trails Claude Opus 5 on HLE w/o tools (56.3% vs 42.6%/44.0† for this release) and on Terminal-Bench 4.0 (51.8% vs 42.3%).
Best for: teams already on DeepSeek-V4.1-Flash wanting a faster, cheaper drop-in agentic/coding upgrade without waiting for a full version bump.
What DeepSeek-V4.1-Flash-1017 Is
DeepSeek-V4.1-Flash-1017 is DeepSeek's dated incremental release on top of DeepSeek-V4.1-Flash (same Causal Encoder-Decoder MoE backbone, 1M-token context, FP4 KV caching, DSpark speculative decoding). Per the vendor, this update focuses on reasoning, coding, agentic, and visual-benchmark gains while shortening task execution time and output length, which the vendor frames as moving the performance-cost Pareto frontier for agent harnesses (Claude Code harness used for Terminal-Bench 4.0; Mini-SWE-V2 harness for DeepSWE v1.1).
No separate technical report was published for this dated release beyond the model card's comparison table; architecture and training details are inherited from the base DeepSeek-V4.1-Flash technical report.
Specifications
| Organization | DeepSeek |
| License | MIT (open weights) |
| Release date | October 17, 2026 (HF) |
| Modality | Multimodal (text + vision) |
| Knowledge cutoff | Not disclosed |
| Context window | 1M tokens |
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| Terminal-Bench 4.0 | 42.3 Pass@1 (Claude Code harness) | DeepSeek model card | Oct 2026 |
| DeepSWE v1.1 | 77.0 Resolved (Mini-SWE-V2 harness) | DeepSeek model card | Oct 2026 |
| ProgramBench | 29.1 Almost@1 | DeepSeek model card | Oct 2026 |
| NL2Repo-Bench | 68.2 | DeepSeek model card | Oct 2026 |
| Agents' Last Exam | 35.4 Pass@1 | DeepSeek model card | Oct 2026 |
| Chartography | 83.0 (with tools) | DeepSeek model card | Oct 2026 |
| BabyVision | 92.5 Pass@1 (with tools) | DeepSeek model card | Oct 2026 |
| ZeroBench | 51.0 Pass@5 (main, with tools) | DeepSeek model card | Oct 2026 |
Scores as reported in DeepSeek's DeepSeek-V4.1-Flash-1017 model card (huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017, October 2026); not independently verified by Benchgen.
Excluded from the table above (bucket-C, documented per scout procedure):
- HLE w/o tools (42.6%, 44.0% text-only subset) — excluded. Our existing
humanitys-last-examleaderboard entries for this model family (e.g. the base V4.1-Flash's 63.9%) were reported with tools, max effort; this release's self-reported number uses a different (no-tools) condition, so it isn't a like-for-like addition to that leaderboard without risking a methodology-mismatch — flagged for manual review rather than guessed. - Automation-Bench v1.0.6 (60.5%) — excluded. No confirmed matching local entity; our existing
automationbenchbenchmark (Zapier) is unversioned and we could not confirm this versioned name is the same benchmark without further research. - Apex Bench v1.1 (72.1%) — excluded. Ambiguous: could map to Mercor's
apex-agentsorapex-swe, or be an unrelated third benchmark; name alone is insufficient to disambiguate, so excluded per the "never silently narrow/widen scope" rule.
Specs and scores sourced from DeepSeek's official model card; third-party benchmark scores attributed inline. Last updated 2026-10-18.