GBench: DeepMind open-sources evaluation suite with 86 benchmark suites
Comments
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Ranked by a public rubric, not by votes: what a comment adds counts, not who wrote it. How comments are ranked · Comment rules
Summary
Google DeepMind has open-sourced gBench, a benchmarking suite for open large language and vision-language models. It measures serving performance on the user's own hardware and capability through three tiers: 86 academic eval suites, 16 deterministic Golden Set smoke tests, and an agentic quality suite.