Benchmarks
Performance and retrieval results
Published runs use pinned corpora, queries, models, and comparison baselines.
Reports
- Released model cost and concurrent latency (repeated runs, p99, memory, CPU, and disk writes)
- Historical model screening (one repetition)
- Latest measured release million-chunk evidence
- Million-chunk methodology and historical paired study
- Public code-retrieval quality
- Public notes and memory retrieval
- Challenge retrieval
- Repository file localization (issue text to fixed files) (rendered on GitHub; results JSON: file-localization-hash-2026-08-21.json)
- Current evidence and release history