MiroThinker-1.7 & H1: Towards Heavy-Duty Research Agents via Verification
Authors: MiroMind Team, S. Bai, L. Bing, L. Lei, R. Li, X. Li, X. Lin, E. Min, L. Su, B. Wang, L. Wang, L. Wang, S. Wang, X. Wang, Y. Zhang, Z. Zhang, G. Chen, L. Chen, Z. Cheng, Y. Deng, Z. Huang, D. Ng, J. Ni, Q. Ren, X. Tang, B. L. Wang, H. Wang, N. Wang, C. Wei, Q. Wu, J. Xia, Y. Xiao, H. Xu, X. Xu, C. Xue, Z. Yang, Z. Yang, F. Ye, H. Ye, J. Yu, C. Zhang, W. Zhang, H. Zhao, P. Zhu
We present MiroThinker-1.7, a new research agent designed for complex long-horizon reasoning tasks. Building on this foundation, we further introduce MiroThinker-H1, which extends the agent with heavy-duty reasoning capabilities for more reliable multi-step problem solving. In particular, MiroThinker-1.7 improves the reliability of each interaction step through an agentic mid-training stage that emphasizes structured planning, contextual reasoning, and tool interaction. This enables more effective multi-step interaction and sustained reasoning across complex tasks. MiroThinker-H1 further incorporates verification directly into the reasoning process at both local and global levels. Intermediate reasoning decisions can be evaluated and refined during inference, while the overall reasoning trajectory is audited to ensure that final answers are supported by coherent chains of evidence. Across benchmarks covering open-web research, scientific reasoning, and financial analysis, MiroThinker-H1 achieves state-of-the-art performance on deep research tasks while maintaining strong results on specialized domains. We also release MiroThinker-1.7 and MiroThinker-1.7-mini as open-source models, providing competitive research-agent capabilities with significantly improved efficiency.
Figures & tables
Figure 1: Comparison of MiroThinker with state-of-the-art agents and agentic foundation models.
Figure 2: The overview of MiroThinker-1.7 & H1.
Figure 3: Overview of the dual-pipeline QA synthesis framework. The Corpus-based Pipeline (left) focuses on topical breadth and high-throughput generation from document subgraphs. The WebHop Pipeline (right) constructs calibrated reasoning trees with web-augmented expansion and hierarchical verification to ensure reasoning rigour and controllable complexity.
Figure 4: The agentic training pipeline of MiroThinker-1.7.
Figure 5: Training dynamics of MiroThinker-1.7-mini for GRPO Agentic RL. BrowseComp-200 is our selected challenging subset from BrowseComp for faster evaluation during training. The plotted curves represent a running average with a window size of 5 to highlight the optimization trends.
Benchmarks
Browse Comp
Browse Comp-ZH
Humanity’s Last Exam
GAIA
xbench-DeepSearch -2510
SEAL-0
DeepSearchQA
Qwen3.5-397B [ 4 ]
78.6
70.3
48.3
–
–
46.9
–
Tongyi-DeepResearch-30B [ 25 ]
43.4
46.7
32.9
70.9
55.0
–
–
GLM-5.0 [ 3 ]
75.9
72.7
50.4
–
–
–
–
Minimax-M2.5 [ 2 ]
76.3
–
–
–
–
–
–
DeepSeek-V3.2 [ 17 ]
67.6
65.0
40.8
–
–
49.5
60.9
Kimi-K2.5 [ 18 ]
78.4
–
50.2
–
46.0
57.4
77.1
Table 1: Performance comparison across various agentic benchmarks. We report the latest publicly available benchmark results for competing models, with the corresponding scores taken from the technical reports or model cards of other organizations. To minimize the impact of randomness from agent-environment interactions on benchmark performance evaluation, we report the average performance of our MiroThinker-1.7 models on each benchmark. We use avg@3 for BrowseComp, BrowseComp-ZH, Humanity’s Last Exam, and DeepSearchQA, and avg@8 for GAIA, xbench-DeepSearch-2510, and SEAL-0.
Model
FrontierSci -Olympiad
SUPERChem (text only)
FinSearchComp (T2/T3)
MedBrowse Comp
Qwen3.5-397B
60.6
49.6
60.8
47.9
Seed-2.0-Pro
74.0
53.0
70.2
–
GPT-5.2-high
77.1
58.0
73.8
–
Claude-4.5-Opus
71.4
43.2
66.2
–
Gemini-3-Pro
76.1
63.2
52.7
–
Kimi-K2.5
–
–
67.8
–
Table 2: Performance comparison across multiple professional-domain benchmarks, including scientific, financial, and medical domains. Some scores are taken from technical reports released by other organizations, while the results for Qwen3.5-397B are obtained through our internal evaluation. To mitigate the randomness introduced by agent–environment interactions during evaluation, we report the mean performance of our MiroThinker-1.7 models on each benchmark. We use avg@3 for FinSearchComp MedBrowseComp, and avg@8 for FrontierSci-Olympiad and SUPERChem.
Model
Report
Factuality
Overall
Grok Deep Research
57.4
58.0
57.7
Manus-1.6-Max Wide Research
53.6
76.4
65.0
Doubao Deep Research
65.8
65.8
65.8
Qwen-3.5-Plus Deep Research
62.4
73.6
68.0
Claude-Opus-4.6 Research
69.9
66.2
68.0
MiniMax-M2.5 Research
62.2
76.4
69.3
Table 3: Long report evaluation on 50 deep research queries automatically generated using the DeepResearchEval query generation framework.
Figure 6: Performance vs. average interaction rounds. Arrows trace improvements from MiroThinker-1.5-30B to MiroThinker-1.7-mini (30B). All trajectories move upper-left, indicating higher performance with fewer turns.
Model
Pass@1
Δ
Steps
Δ
MiroThinker-1.7
32.1
–
1185.2
–
MiroThinker-H1 w/ Local Verifier Only
58.5
+26.4
210.8
-974.4
Table 4: Local Verification only on BrowseComp hard subset (295 questions). Steps include all retries.
Figure 7: Token scaling curve of MiroThinker-H1 on BrowseComp. At 16× compute, the default budget for all benchmarks, accuracy reaches 85.9. Scaling to 64× further improves accuracy to 88.2.
Autonomous research agents use language models and tools to search literature, design and run experiments, analyze results, and support scientific discovery.