DeepSeek Releases V4.1-Flash-1017 Open Weights Under MIT; Artificial Analysis Scores It 45.10
DeepSeek released DeepSeek-V4.1-Flash-1017 on Oct. 17 as MIT-licensed weights, a 763B-parameter model with 1M-token context that it calls a minor update. Artificial Analysis scores it 45.10 on its Intelligence Index, second among open-weight models in the payload we parsed behind MiMo-V2.6-Pro at 46.32. Lab agentic gains are DeepSeek's own.
A cheap, permissively licensed Chinese open model moved up on agentic coding the same week a DeepSeek Flash variant appeared in a hacking campaign and DHS ordered DeepSeek out of civilian agencies. DeepSeek released DeepSeek-V4.1-Flash-1017 on Saturday, October 17 as MIT-licensed weights on Hugging Face and ModelScope. The model card calls it a "minor update" to a 763B-parameter model with a 1M-token context.
Lab-reported scores#
Compared with base V4.1-Flash, DeepSeek reports Humanity's Last Exam without tools at 42.6 vs 36.8; Terminal-Bench 4.0 at 42.3 vs 31.2 (Claude Code harness; the card lists Opus 5 at 51.8); DeepSWE v1.1 at 77.0 vs 74.2 (Mini-SWE-V2 harness); and ProgramBench at 29.1 vs 20.3. The Decoder says DeepSeek also opened Search and Fetch APIs at $0.75 per 1,000 calls and that DeepSWE tasks cost just under $0.20 each against about $0.50 for the predecessor.
Independent score#
Artificial Analysis scores DeepSeek V4.1 Flash 1017 (Max) at 45.10 on its Intelligence Index. In the payload we parsed, that is second among open-weight models behind MiMo-V2.6-Pro (46.32). We covered earlier V4 pricing separately. NorForge's Cairn worker used an unspecified deepseek-v4.1-flash snapshot; we do not say 1017 was used.
Limits#
- Agentic and coding figures are DeepSeek's, run in different harnesses. Only the Artificial Analysis index is independent here.