DeepSeek 4.1-1017 benchmarks leak: roughly matches Opus 5, beats GLM5.3 across the board
kimmonismus · x · 2026-10-17
Blogger kimmonismus cites a new DeepSeek 4.1-1017 (DS-V4.1-Flash-1017) benchmark table, claiming it's roughly on par with Opus 5 and a major jump over the regular 4.1 Flash.
Key numbers from the image:
- HLE w/o tools: 42.6 (44.0 with tools), up from 36.8
- Terminal-Bench 4.0: 42.3, ahead of GLM5.3 (37.9) and GPT-5.6 Sol (39.9), behind Opus 5 (51.8)
- DeepSWE v1.1: 77.0 — best in the table, edging Opus 5 (74.0)
- Agents' Last Exam: 35.4, table-best
- Outperforms GLM5.3 and GPT-5.6 Sol on most rows (ProgramBench, NL2Repo-Bench, Apex Bench, Chartography, BabyVision), though Opus 5 still leads on several
The author suggests buying two DGX Spark units to run it locally. Note: third-party-shared scores; official leaderboards and hands-on results pending.
More from Models
- Unverified: ChatGPT GPT-6 Pro now splits tasks across parallel agents for Pro users — mark_k · 2026-10-17
- Distillation angle overrated: frontier reasoning gains mostly come from internal RL, argues analyst — teortaxesTex · 2026-10-17
- Meta's Muse AI stumbles: needs login and virtual browser just to fetch a Marketplace photo — robblatt · 2026-10-17
- Ultra subscriber: Gemini 4 Argon weekly quotas so low they run out in two days — DeryaTR_ · 2026-10-17
- Codex upgrades DSH and live-tests DeepSeek-V4.1-Flash 1017 — cedric_chee · 2026-10-17
- Will Human Brains Soon Fail to Judge Whether New Models Are Actually Better? — arjunrajlab · 2026-10-17