Benchmarks · Overview
Every score. Every run. Numbers you can verify.
Bahulam Code is measured on the standard coding-agent benchmarks — SWE-bench Lite, Verified, and Full, plus Terminal-bench. Every run is published with full predictions, official evaluator output, and costs in our public benchmarks repository.
Last updated: 2026-07-30
Latest — SWE-bench Lite 2.6.14
58.7%
176 / 300 resolved · MiMo v2.5 · 95% cache · — total
SWE-bench
Lite · Verified · Full
Same task shape (auto-generated GitHub-issue → patch). Lite = 300 hand-selected issues, Verified = the reproducible subset, Full = the entire 2 294. We publish deep-views for each.
SWE-bench LitePublic repo
bahulam-benchmarks on GitHub
Every run in this section is backed by a public artifact bundle: predictions.json, official Docker-eval report, cost breakdown, and the exact config. Clone it, verify it, rerun it.
Browse the repoComparison
Bahulam Code across models
Same agent. Same harness. Only the model changes. One column per model: MiMo v2.5, DeepSeek Flash, GLM 5.2, Tencent HY3. Compare $/resolved and cache-hit rate side-by-side.
Compare modelsReference
Methodology
What SWE-bench measures, what "resolved" means (Docker-eval spec), how the Bahulam Code harness invokes the CLI. Read this before challenging the numbers.
Read the methodology