Atlas

Benchmarks

← All benchmarks

PinchBench (Meta container variant)

Agents · 2026-10-17

Meta's container-based variant of PinchBench in the Muse Glimmer report: 24 knowledge-work tasks in an isolated scientific-Python sandbox with bash, file, code-execution, web-search and web-fetch tools. Each task combines PinchBench's automated checks with a gpt-oss-120b judge of the deliverable's groundedness, and the score is the mean task reward. Its harness and judge differ from the public PinchBench leaderboard, and the report itself treats the 24-task result as diagnostic.

Top models (higher is better)

ModelScore
Muse Glimmer 30B90.9
Loading Atlas data…