Deep Search on GAIA
81.9AccuracyMiroThinker-v1.0-72B
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MiroThinker-v1.0-72BType=SA2026.02 | 81.9 | — | |
| TaSType=MA, Backbone=GPT-5 Medium-Think2026.02 | 77.7 | — | |
| GPT-5 High-ThinkType=N/A2026.02 | 76.4 | — | |
| MiroThinker-v1.0-30BType=SA2026.02 | 73.5 | — | |
| GPT-5 Medium-ThinkType=MA2026.02 | 71.8 | — | |
| Tongyi-DeepResearch + DS-MCMModel=Tongyi-DeepResearch, System Category=Open-Source, Search Strategy=+ DS-MCM2026.01 | 71.7 | — | |
| Tongyi DeepResearch (30B)Type=SA2026.02 | 70.9 | — | |
| OpenAI o3Model=OpenAI o3, System Category=Proprietary, Search Strategy=Vanilla2026.01 | 70.9 | — | |
| Qwen3-30B-A3B-thinking-SFT + SAPOContext Budget=128K2026.05 | 70.9 | — | |
| OpenAI DeepSearchModel=OpenAI DeepSearch, System Category=Proprietary, Search Strategy=Vanilla2026.01 | 70.5 | — | |
| OpenAI-o3Context Budget=-2026.05 | 70.5 | — | |
| Tongyi-DeepResearchModel=Tongyi-DeepResearch, System Category=Open-Source, Search Strategy=Vanilla2026.01 | 70.1 | — | |
| Tongyi-DeepResearch + LLM-CriticModel=Tongyi-DeepResearch, System Category=Open-Source, Search Strategy=+ LLM-Critic2026.01 | 69.3 | — | |
| MiroThinker-DeepResearch + DS-MCMModel=MiroThinker-DeepResearch, System Category=Open-Source, Search Strategy=+ DS-MCM2026.01 | 69.3 | — | |
| Tongyi DeepResearchContext Budget=128K2026.05 | 68.9 | — | |
| Claude-4-Sonnet (Thinking)Type=SA2026.02 | 68.3 | — | |
| OpenAI Deep ResearchType=N/A2026.02 | 67.4 | — | |
| MiroThinker-v1.0-8BType=SA2026.02 | 66.4 | — | |
| GPT-5 Medium-ThinkType=SA2026.02 | 66 | — | |
| MiroThinker-DeepResearchModel=MiroThinker-DeepResearch, System Category=Open-Source, Search Strategy=Vanilla2026.01 | 63.8 | — | |
| WebSailor-32BContext Budget=32K2026.05 | 63.1 | — | |
| MiroThinker-DeepResearch + LLM-CriticModel=MiroThinker-DeepResearch, System Category=Open-Source, Search Strategy=+ LLM-Critic2026.01 | 63 | — | |
| OpenAI GPT-5Model=OpenAI GPT-5, System Category=Proprietary, Search Strategy=Vanilla2026.01 | 62.8 | — | |
| Qwen3-30B-A3B-thinking-SFT + GRPOContext Budget=128K2026.05 | 62.1 | — | |
| Gemini 2.5 ProModel=Gemini 2.5 Pro, System Category=Proprietary, Search Strategy=Vanilla2026.01 | 60.6 | — | |
| Gemini-2.5-ProType=SA2026.02 | 60.2 | — | |
| Kimi K2Context Budget=-2026.05 | 57.7 | — | |
| Qwen3-30B-A3B-thinking-SFTContext Budget=128K2026.05 | 57.3 | — | |
| TaSType=MA, Backbone=Gemini-2.5-Flash2026.02 | 52.4 | — | |
| Qwen3-8B-SFT + SAPOContext Budget=64K2026.05 | 52.4 | — | |
| Qwen3-MaxType=MA2026.02 | 52 | — | |
| Qwen3-8B-SFT + GRPOContext Budget=64K2026.05 | 51.5 | — | |
| WebExplorer-8BContext Budget=64K2026.05 | 50.5 | — | |
| TaSType=MA, Backbone=Qwen3-Max2026.02 | 49 | — | |
| Qwen3-8B-SFT + ARPOContext Budget=64K2026.05 | 47.3 | — | |
| Qwen3-30B-MoE + DS-MCMModel=Qwen3-30B-MoE, System Category=Open-Source, Search Strategy=+ DS-MCM2026.01 | 47.2 | — | |
| Grok-3 DeeperSearchModel=Grok-3 DeeperSearch, System Category=Proprietary, Search Strategy=Vanilla2026.01 | 47.1 | — | |
| Web-30B-E-GRPOContext Budget=-2026.05 | 46.7 | — | |
| Qwen3-30B-MoE + LLM-CriticModel=Qwen3-30B-MoE, System Category=Open-Source, Search Strategy=+ LLM-Critic2026.01 | 43.3 | — | |
| Qwen3-30B-MoEModel=Qwen3-30B-MoE, System Category=Open-Source, Search Strategy=Vanilla2026.01 | 40.9 | — | |
| Qwen3-MaxType=SA2026.02 | 39.8 | — | |
| Qwen3-8B-SFTContext Budget=64K2026.05 | 38.8 | — | |
| Gemini-2.5-FlashType=MA2026.02 | 38.4 | — | |
| WebSailor-7BContext Budget=32K2026.05 | 35.9 | — | |
| MiroThinker-v1.0-8BContext Budget=64K2026.05 | 31.2 | — | |
| SAGEBackbone=QWEN-7B, Training data=SAGE (Ours), Search Engine=Google Search, Max queries=102026.01 | 24 | — | |
| MiroThinker-v1.5-30BContext Budget=128K2026.05 | 23.3 | — | |
| SAGEBackbone=QWEN-3B, Training data=SAGE (Ours), Search Engine=Google Search, Max queries=102026.01 | 18.8 | — | |
| Gemini-2.5-FlashType=SA2026.02 | 16.3 | — | |
| MusiqueBackbone=QWEN-7B, Training data=Musique, Search Engine=Google Search, Max queries=102026.01 | 15.6 | — | |
| RFTBackbone=Qwen3-8B2026.04 | 14.98 | — | |
| NQ + HotpotQABackbone=QWEN-7B, Training data=NQ + HotpotQA, Search Engine=Google Search, Max queries=102026.01 | 14.6 | — | |
| ProCeedRLBackbone=Qwen3-8B2026.04 | 13.79 | — | |
| MusiqueBackbone=QWEN-3B, Training data=Musique, Search Engine=Google Search, Max queries=102026.01 | 13.5 | — | |
| NQ + HotpotQABackbone=QWEN-3B, Training data=NQ + HotpotQA, Search Engine=Google Search, Max queries=102026.01 | 12.5 | — | |
| Qwen3-8B-v3-SFTBackbone=Qwen3-8B2026.04 | 10.51 | — | |
| DAPO/Search-R1Backbone=Qwen3-8B2026.04 | 10.1 | — | |
| Rewinding MoreBackbone=Qwen3-8B, Type=ablation2026.04 | 10.1 | — | |
| ReAct PromptingBackbone=Qwen3-8B2026.04 | 9.69 | — |