Skip to main content





Qwen3.7-Max Benchmark Scores
Alibaba’s Qwen team has introduced Qwen3.7-Max, a new frontier AI model focused heavily on coding agents, autonomous execution, and long-horizon reasoning
Last updated: May 20, 2026
Alibaba’s Qwen team has introduced Qwen3.7-Max, a new frontier AI model focused heavily on coding agents, autonomous execution, and long-horizon reasoning. The release is structured almost entirely around benchmark performance and sustained agent behavior rather than chatbot interaction, making it one of the clearest examples yet of how frontier AI labs are shifting toward autonomous systems that can operate inside software environments for extended periods of time.
The company directly compares Qwen3.7-Max against several major frontier systems including Opus-4.6 Max, DeepSeek V4 Pro Max, GLM-5.1 Thinking, and Kimi K2.6 Thinking. Across the benchmark table, Qwen3.7-Max consistently lands near the top and in several cases takes the leading position outright.
What stands out most is not necessarily one single benchmark win, but the breadth of performance. The model performs strongly across coding, software engineering, tool use, multilingual reasoning, long-context evaluation, and difficult STEM tasks simultaneously. Alibaba is clearly trying to position Qwen3.7-Max as a general-purpose autonomous agent foundation rather than a narrowly optimized coding model.

Coding Benchmark Performance
The coding benchmarks are among the strongest parts of the release and show Alibaba aggressively targeting practical software-engineering workloads rather than traditional algorithmic coding tests.

On Terminal Bench 2.0-Terminus, Qwen3.7-Max scored 69.7, outperforming DeepSeek V4 Pro Max at 67.9 and Opus-4.6 Max at 65.4. This benchmark evaluates models operating directly inside terminal environments where they must issue commands, debug failures, and interact with software systems over long sessions. The result suggests Qwen3.7-Max is particularly effective at sustained coding-agent workflows rather than isolated code generation.
The model also performed extremely well across the SWE benchmark family, which is increasingly viewed as one of the best measurements of real software-engineering ability. On SWE-Pro, Qwen3.7-Max reached 60.6, slightly ahead of Kimi K2.6 Thinking at 59.5 and DeepSeek V4 Pro Max at 59.0. On SWE-Multilingual, the model achieved 78.3, beating Opus-4.6 Max at 77.5 and DeepSeek V4 Pro Max at 76.2.
Perhaps most importantly, Qwen3.7-Max scored 80.4 on SWE-Verified, effectively tying Opus-4.6 Max at 80.8 and DeepSeek V4 Pro Max at 80.6. That places the model firmly in the highest tier of current software-engineering systems.
The release also highlights strong repository-level and frontend-generation performance. Qwen3.7-Max scored 47.2 on NL2Repo and 53.5 on SciCode, outperforming Opus-4.6 Max’s 51.9 result on the latter benchmark. On frontend-focused evaluations, the model achieved 1568 on QwenWebDev and 1608 on QwenSVG, narrowly surpassing GLM-5.1 Thinking’s 1605 score on SVG generation tasks.
These benchmarks matter because they evaluate increasingly important categories of AI coding behavior: repository understanding, scientific programming, frontend application generation, and graphical interface construction. Alibaba is clearly emphasizing that Qwen3.7-Max can function across many different categories of engineering work rather than excelling only at narrow benchmark tasks.
The model does not dominate every single evaluation. DeepSeek V4 Pro Max still leads LiveCodeBench with 93.5 compared to Qwen3.7-Max’s 91.6. But across the full coding benchmark suite, Qwen3.7-Max appears unusually balanced and consistently competitive.
Agent and Tool-Use Results
The strongest relative gains for Qwen3.7-Max may actually appear in the agent and tool-use category, which is increasingly becoming the most strategically important area of frontier AI development.


On MCP-Mark, the model scored 60.8, outperforming GLM-5.1 Thinking at 57.5 and Opus-4.6 Max at 56.7. MCP benchmarks evaluate how effectively models interact with external tools and APIs through structured orchestration systems. These capabilities are becoming central to enterprise AI deployment because modern agents increasingly rely on browsers, terminals, retrieval systems, APIs, and external execution environments.
Qwen3.7-Max also led MCP-Atlas with a score of 76.4, narrowly ahead of Opus-4.6 Max at 75.8. On SkillsBench, the model achieved 59.2, beating Kimi K2.6 Thinking at 56.2 and DeepSeek V4 Pro Max at 52.3.
Function-calling reliability also remained strong. On BFCL-V4, Qwen3.7-Max reached 75.0, placing it very close to Opus-4.6 Max at 76.7. Reliable function calling is one of the most important capabilities for autonomous agents because execution errors compound rapidly during long workflows.
The release also highlighted office-style and structured workflow tasks. On SpreadSheetBench-v1, Qwen3.7-Max achieved 87.0, placing it among the strongest systems currently tested for spreadsheet reasoning and workflow execution. Spreadsheet benchmarks are increasingly important because they approximate real-world business tasks more closely than conversational evaluations.
Another important result came from QwenWorldBench, where the model scored 57.3. Alibaba describes this benchmark as testing models acting as “world models” across environments involving terminals, browsers, Android systems, operating systems, and software-engineering tasks. That reflects the industry-wide shift toward evaluating models as autonomous actors operating inside dynamic environments rather than static text predictors.
Reasoning and STEM Performance
Qwen3.7-Max also posted some of the strongest reasoning scores in the release, suggesting Alibaba has not sacrificed raw reasoning capability in pursuit of better agent behavior.


On GPQA Diamond, one of the hardest scientific reasoning benchmarks currently used by frontier labs, the model scored 92.4. That surpassed Opus-4.6 Max at 91.3 and DeepSeek V4 Pro Max at 90.1.
On HLE, which evaluates difficult long-form reasoning tasks, Qwen3.7-Max reached 41.4 compared to Opus-4.6 Max’s 40.0. The model also achieved 97.1 on HMMT 2026 Feb and 90.0 on IMOAnswerBench, placing it among the strongest publicly discussed systems for mathematical reasoning.
One of the most dramatic results appeared on Apex, where Qwen3.7-Max scored 44.5. DeepSeek V4 Pro Max reached 38.3 while Opus-4.6 Max scored 34.5. The size of that gap suggests Alibaba sees advanced reasoning as one of Qwen3.7-Max’s major strengths rather than simply a supporting capability.
The benchmark table overall paints a picture of a model that combines frontier-level reasoning with unusually strong execution ability across software environments.
The 35-Hour Autonomous Kernel Optimization Test
The most technically ambitious section of the release involved a 35-hour autonomous optimization run on previously unseen hardware.
Alibaba tasked Qwen3.7-Max with optimizing an Extend Attention kernel on a T-Head ZW-M890 PPU architecture that the model had allegedly never encountered during training. According to the company, the system had no hardware documentation, no profiling data, and no architecture-specific examples available.
Over approximately 35 hours, Qwen3.7-Max performed 432 kernel evaluations and issued 1,158 tool calls while iteratively redesigning and optimizing the kernel. The final reported result was a 10x speedup over the baseline Triton implementation.
Alibaba compared the same task against several competing models under identical conditions. GLM 5.1 reportedly achieved 7.3x speedup, Kimi K2.6 reached 5.0x, DeepSeek V4 Pro reached 3.3x, and Qwen3.6-Plus only managed 1.1x.
The significance of the experiment is less about raw benchmark numbers and more about persistence. Alibaba is arguing that Qwen3.7-Max can sustain coherent reasoning and productive optimization over extremely long execution windows involving hundreds or thousands of actions without collapsing into repetitive loops or losing strategic direction. That kind of long-horizon execution is increasingly viewed as one of the defining challenges for autonomous AI systems.
What Makes The Release Important
The release matters because it reflects a major transition happening across frontier AI development. The focus is no longer primarily on chatbot interaction or static prompt-response behavior. Nearly every benchmark emphasized by Alibaba involves autonomous execution, tool use, repository editing, long-context reasoning, workflow management, or environment interaction.
Qwen3.7-Max is being positioned less as a conversational assistant and more as a foundational operating system for autonomous software agents. The benchmark results also show how competitive Chinese frontier labs have become. Qwen3.7-Max is now trading benchmark wins directly with some of the strongest proprietary AI systems currently available.
Conclusion
Qwen3.7-Max is one of the strongest autonomous-agent-focused AI releases so far this year. The model performs particularly well across coding-agent evaluations, software-engineering benchmarks, tool-use tasks, and advanced reasoning tests. In several important categories, including Terminal Bench 2.0, GPQA Diamond, MCP-Mark, and Apex, it either leads the field outright or lands extremely close to the top.
The central claim of the release is sustained autonomy. Alibaba is effectively arguing that Qwen3.7-Max can remain productive across extremely long execution windows involving hundreds or thousands of actions while maintaining coherent reasoning and strategic consistency.
Whether every claim fully survives broader independent testing remains unclear, but the benchmark results alone place Qwen3.7-Max firmly among the leading frontier AI systems currently available.
Add a comment
Tags: ML News