Alibaba Unveils Qwen3.7-Max: Autonomous AI Agent Achieves 10x Speedup in 35-Hour Kernel Optimization
On May 20, 2026, Alibaba officially launched Qwen3.7-Max at the Alibaba Cloud Summit.
Overview
Qwen3.7-Max is Alibaba's latest flagship model, deeply optimized for agent scenarios. It delivers top-tier performance in programming, reasoning, office automation, and long-duration task execution, achieving comprehensive capabilities on par with leading international models like GPT, Claude, and Gemini.
Key Benchmark Results
Programming Agent
| Benchmark | Qwen3.7-Max | DeepSeek-v4-Pro Max | Claude Opus 4.7 Max |
|---|---|---|---|
| TerminalBench 2.0-Terminus | 69.7 | 67.9 | 65.4 |
| SWE-Multilingual | 78.3 | - | - |
| SWE-Pro | 60.6 | - | - |
General Agent
| Benchmark | Qwen3.7-Max | Claude Opus 4.6 | GLM 5.1 |
|---|---|---|---|
| MCP-Atlas | 76.4 | 75.8 | - |
| MCP-Mark | 60.8 | - | 57.5 |
| SpreadSheetBench-v1 | 87.0 | - | - |
Reasoning Ability
| Benchmark | Qwen3.7-Max | Claude Opus 4.6 |
|---|---|---|
| GPQA Diamond | 92.4 | 91.3 |
| HLE | 41.4 | 40.0 |
35-Hour Autonomous Kernel Optimization Experiment
The most notable achievement of Qwen3.7-Max is its fully autonomous hardware optimization task lasting 35 hours.
Alibaba tasked Qwen3.7-Max with optimizing inference kernels on an unseen chip (T-Head ZhenWu M890) not present in training data. Without human intervention, the model worked continuously for 35 hours, ultimately improving the performance of Triton operators by 10x.
Experiment Details
- Chip: T-Head ZhenWu M890 (no training data exposure)
- Work Duration: 35 hours continuous
- Tool Invocation Count: 1,158 times
- Kernel Evaluation Count: 432 times
- Final Result: 10x performance improvement over official reference implementation
Comparison with Other Models
| Model | Geometric Mean Speedup |
|---|---|
| Qwen3.7-Max | 10.0x |
| GLM 5.1 | 7.3x |
| Kimi K2.6 | 5.0x |
| DeepSeek V4 Pro | 3.3x (interrupted midway) |
Artificial Analysis Ranking
According to the latest ranking from third-party evaluator Artificial Analysis:
- Overall Score: 56.6 points
- Global Ranking: 5th place
- Domestic Models: 1st place
- Progress from Previous Generation: +4.8 points
The top positions include models like GPT-5.4(xhigh), Gemini 3.1 Pro Preview, and Claude-Opus4.7(max).
Release Pace
The Qwen series maintains a rapid iteration pace:
| Date | Model | Theme |
|---|---|---|
| March 20, 2026 | Qwen3.5-Max-Preview | Toward Native Multimodal Agents |
| April 20, 2026 | Qwen3.6-Max-Preview | Toward Real-World Agents |
| May 20, 2026 | Qwen3.7-Max | New Benchmark for the Agent Era |
Each month, a flagship model is released, consistently pushing the performance ceiling for domestic models.
Conclusion
Qwen3.7-Max represents a new-generation flagship model specialized in agent capabilities, achieving top-tier performance across programming, reasoning, and office automation. Particularly, the 35-hour autonomous kernel optimization experiment marks a significant milestone in demonstrating AI models' long-term autonomous operational abilities.
Loading...