Agents' Last Exam benchmark runs fully on HF infra; Xiaomi MiMo-V2.6 tops leaderboard at 31.6
yb2698 · x · 2026-10-21
Agents' Last Exam, an agent benchmark led by Berkeley RDI with 300+ industry experts, now runs 100% on Hugging Face infrastructure and has been used by OpenAI, Anthropic, Qwen, DeepSeek and more for model releases.
- Infra consolidation: agent benchmarks traditionally need a model server, per-task sandboxes, result storage, and glue scripts across 2-3 clouds; now HF Jobs handles servers/sandboxes/orchestration, HF Buckets provide cross-job data sharing, and traces render automatically — one token, no glue code.
- Benchmark scope: 1,500+ tasks collected toward a 5,000-task target across all 55 sub-industries, testing long-horizon, economically valuable work in After Effects, Siemens NX, Unreal Engine, manufacturing simulation, neuroimaging and more.
- Leaderboard (screenshot): Xiaomi MiMo-V2.6-Pro-RL leads at 31.6, Kimi K3 second at 28.29, MiMo-V2.6-Flash-RL third at 27.6, Qwen3.8-Flash-Next fifth at 24.3 — Chinese models dominate the top spots.
More from coding & agent
- Three ways to let Claude Opus 5.5 click around your computer: cua-driver, OpenAI CUA MCP, or OSWorld harness — zsakib_ · 2026-10-21
- REUF2F Benchmarks Coding Agents on Formalized Undergraduate Research Problems — wellecks · 2026-10-21
- At NVIDIA GTC Berlin, OpenClaw Drives a Full Robotic Simulation on a Single DGX Spark — vincent_koc · 2026-10-21
- US healthtech deploys AI agents to collect patient bills, paid only on recovery — alex_verem · 2026-10-21
- Where should market agents stop? Default to monitoring, require approval to act — Decent_Wing_4393 · 2026-10-21
- Testing Codex × CapCut Web: turning a brand campaign brief into video via natural language — Div_pradeep · 2026-10-21