FRAMES (Google multi-hop retrieval+reasoning QA) — Distinct corpus (13 models/36 rows, DeepSeek-sourced) vs Kimi-card Frames.
| # | Model | Vendor | Best score | Runs | Last seen |
|---|---|---|---|---|---|
| 1 | DeepSeek-R1 | DeepSeek | 83 | 10 | 2026-06-04 |
| 2 | R1-Zero | DeepSeek | 82.3 | 1 | 2026-05-03 |
| 3 | GPT-4o | OpenAI | 80.5 | 6 | 2026-06-04 |
| 4 | OpenAI o1-mini | OpenAI | 76.9 | 5 | 2026-06-04 |
| 5 | DeepSeek-V3 | DeepSeek | 73.3 | 6 | 2026-06-04 |
| 6 | Claude 3.5 Sonnet | Anthropic | 72.5 | 6 | 2026-06-04 |
| 7 | Llama 3.1 Instruct 405B | Meta | 70 | 1 | 2026-05-03 |
| 8 | Qwen2.5 Instruct 72B | Alibaba | 69.8 | 1 | 2026-05-03 |
| 9 | DeepSeek-V2 | DeepSeek | 66.9 | 1 | 2026-05-03 |
| 10 | DeepSeek-V2.5 | DeepSeek | 65.4 | 1 | 2026-05-03 |
Best tracked score per model (default configuration; source-attributed and verification-tiered). Open a model for its full benchmark surface, provenance and pricing.