MiroThinker-1.7 & H1: Open-Source Research Agents That Beat Frontier Models
MiroMind AI released MiroThinker-1.7 and MiroThinker-H1 on March 16, 2026 — open-source deep research agents that achieve state-of-the-art results on complex research and prediction benchmarks, outperforming Claude 4.6 Opus, GPT-5.4, and Gemini 3.1 Pro.
MiroThinker-H1 scores 88.2 on BrowseComp (vs. Claude 4.6 Opus at 84.0 and GPT-5.4 at 82.7) and 84.4 on BrowseComp-ZH, leading all evaluated frontier models in both English and Chinese research tasks. On FrontierScience-Olympiad, H1 scores 79.0. The system supports 256K context windows and 300-600 tool interactions per task.
The key innovation is a dual-layer verification system integrated into the model's reasoning process. A Local Verifier audits intermediate reasoning decisions in real time, enabling long-chain verifiable reasoning — each step is both locally and globally verifiable. Models range from 8B to 235B parameters.
The project has 7.1K GitHub stars and is fully open-source with code for both SFT and DPO training. It received 116 upvotes on HuggingFace Daily Papers today.
https://github.com/MiroMindAI/MiroThinker
← Back to all articles
MiroThinker-H1 scores 88.2 on BrowseComp (vs. Claude 4.6 Opus at 84.0 and GPT-5.4 at 82.7) and 84.4 on BrowseComp-ZH, leading all evaluated frontier models in both English and Chinese research tasks. On FrontierScience-Olympiad, H1 scores 79.0. The system supports 256K context windows and 300-600 tool interactions per task.
The key innovation is a dual-layer verification system integrated into the model's reasoning process. A Local Verifier audits intermediate reasoning decisions in real time, enabling long-chain verifiable reasoning — each step is both locally and globally verifiable. Models range from 8B to 235B parameters.
The project has 7.1K GitHub stars and is fully open-source with code for both SFT and DPO training. It received 116 upvotes on HuggingFace Daily Papers today.
https://github.com/MiroMindAI/MiroThinker