MRCR (Multi-Round Co-reference Resolution) — Unscoped/unversioned reading, MiMo source; kept apart from 1M/v2 variants.
| # | Model | Vendor | Best score | Runs | Last seen |
|---|---|---|---|---|---|
| 1 | Gemini 2.5 Pro | 93 | 1 | 2026-08-23 | |
| 2 | Gemini 3 Pro | 89.7 | 2 | 2026-06-13 | |
| 3 | Gemini 1.5 Pro | 82.6 | 1 | 2026-08-23 | |
| 4 | Gemini 2.0 Flash (Reasoning) | 69.2 | 1 | 2026-08-23 | |
| 5 | DeepSeek-V3.2 | DeepSeek | 55.5 | 2 | 2026-06-13 |
| 6 | Claude Sonnet 4.5 | Anthropic | 55.4 | 1 | 2026-06-06 |
| 7 | Gemini 1.5 Flash 8B | 54.7 | 1 | 2026-08-23 | |
| 8 | MiMo V2 Flash | Xiaomi | 45.7 | 3 | 2026-08-23 |
| 9 | Kimi K2 (Reasoning) | Moonshot | 44.2 | 2 | 2026-06-13 |
| 10 | Gemini 2.5 Flash | 32 | 1 | 2026-08-23 |
Best tracked score per model (default configuration; source-attributed and verification-tiered). Open a model for its full benchmark surface, provenance and pricing.