gBench
Compare model capabilities, serving performance and readiness for agent workflows in one evaluation harness.
Explore the original evidence
Explore the original evaluation, its methods, and reported findings.
Open evalCoverage
The current harness combines 86 academic suites, 16 deterministic smoke tests and a separate GemmaClaw agent suite. Luciano Martins and Frank Li are credited as authors.
Methodology
Measure the three capability tiers separately from serving latency and throughput. The agent suite follows a pinned GemmaClaw revision; scenarios requiring unsupported providers are excluded, and harness failures are not model failures. A harness and suite collection, not 86 newly created benchmarks. Underlying datasets retain their own licenses. Software revisions, scaffolds, reasoning settings and hardware affect comparability; no results were rerun.