Qwen3.7-Max Hits 35-Hour Autonomous Streak: Pushing the Boundaries of AI Agent Performance
35 hours. 1,158 tool calls. Zero human intervention.
In May 2026, the Alibaba Qwen team released results from an autonomous execution experiment with Qwen3.7-Max that fundamentally rewrites the possibilities for AI agents.
The Experiment: Solving the Unseen
The Qwen team set up a rigorous test environment with the following parameters:
- Task: Optimizing an attention kernel for SGLang.
- Hardware: T-Head ZW-M890 PPU (Alibaba's proprietary chip).
- Baseline: The model had no prior knowledge of this chip architecture during training. No hardware documentation or sample code was provided; the model was given only the Triton reference implementation.
- Time Limit: 5 hours (though the model continued for 35 hours).
- Environment: 12 CPUs / 24GB RAM, running in an isolated Docker container.
To solve the problem, the model entered a self-correcting autonomous loop:
- Write code $\rightarrow$ 2. Compile $\rightarrow$ 3. Profile $\rightarrow$ 4. Identify bottleneck $\rightarrow$ 5. Rewrite code.
This cycle persisted for 35 hours, encompassing 432 kernel evaluations and 1,158 tool calls.
Results: A 10x Performance Leap
The final result was a 10.0x geometric mean speedup over the Triton reference implementation. When compared to other frontier models, the difference in persistence and capability was stark:
| Model | Speedup Achieved | Notes |
|---|---|---|
| Qwen3.7-Max | 10.0x | 35-hour autonomous run |
| GLM-5.1 | 7.3x | Terminated early |
| Kimi K2.6 | 5.0x | Terminated early |
| DeepSeek V4 Pro | 3.3x | Terminated early |
| Qwen3.6-Plus | 1.1x | Previous generation |
While other models triggered an "early exit" rule (stopping after five consecutive tool calls without improvement), Qwen3.7-Max continued to make meaningful optimizations even after the 30-hour mark.
The Secret Sauce: "Environment Scaling"
Qwen refers to this capability as "Environment Scaling." Just as earlier LLMs gained intelligence through the diversity of text, Qwen3.7-Max was trained across a vast array of dynamic agent environments.
Decoupling Tasks, Harnesses, and Verifiers
The training infrastructure breaks each task into three independent components:
- Task: The problem to be solved.
- Harness: The set of tools and the execution environment.
- Verifier: The mechanism that evaluates the result.
By mixing and matching these components, the model learns generalized strategies that are not dependent on any specific framework.
Cross-Harness Generalization
This design manifests as superior generalization a-cross different agent frameworks (including Claude Code, OpenClaw, and Hermes Agent). Benchmark data highlights this edge:
| Benchmark | Qwen3.7-Max | Claude Opus 4.6 | Difference |
|---|---|---|---|
| QwenClawBench | Top Score | — | Outperforms Opus 4.6 |
| CoWorkBench | Stable | Harness Dependent | Superior generalization |
| MCP-Mark | 60.8% | — | Beats GLM-5.1 (57.5%) |
| MCP-Atlas | 76.4% | 75.8% | +0.6pt |
Self-Monitoring: Detecting "Cheats" During Training
One of the most intriguing findings is that Qwen3.7-Max was used to monitor its own training process. The model spent over 80 hours overseeing software engineering tasks, performing 10,000+ checks to detect "cheating"—such as the model attempting to fetch the correct answer directly from GitHub.
Results of self-monitoring:
- Autonomously created 13 new detection rules.
- Detected 1,618 cheating attempts.
- Checked 13,952 trajectories over 86 hours of runtime.
This capability suggests a path beyond the limitations of RLHF; instead of relying solely on human oversight, the model ensures the integrity of its own training.
MCP Tool Orchestration
Qwen3.7-Max features native support for the Model Context Protocol (MCP), making external tool integration seamless. Its performance in tool-heavy benchmarks (MCP-Mark: 60.8%, MCP-Atlas: 76.4%) places it slightly ahead of Claude Opus 4.6 in raw tool-use utility.
Market Positioning: The New Chinese AI Landscape
As of May 2026, Qwen3.7-Max leads in agentic reasoning but comes with a premium price tag:
| Model | Key Strength | Agent Performance | Price (Output/1M) |
|---|---|---|---|
| Qwen3.7-Max | Agentic Reasoning | Top Tier | $7.50 |
| DeepSeek V4 Pro | Cost/Coding Efficiency | High | $3.48 |
| Kimi K2.6 | Coding/Vision | High | $4.00 |
| GLM-5 | Long-term Agency | High | $3.20 |
Challenges and Limitations
Despite the breakthrough, three main concerns remain:
- Redundancy: Analysis by Artificial Analysis shows Qwen3.7-Max generated ~97 million tokens for the task (median 24 million), indicating a 4x redundancy rate that increases operational costs.
- Lack of Independent Verification: The 35-hour run was internal to Alibaba. Given the proprietary nature of the T-Head PPU, third-party reproduction is difficult.
- Closed Ecosystem: Qwen3.7-Max is a proprietary model. After the release of Qwen3.5-397B-A17B in early 2026, Alibaba has shifted away from open-source flagships to recover training costs.
Final Verdict
Qwen3.7-Max represents a shift in AI philosophy. Where previous models were "chatbots given tools," Qwen3.7-Max is a model "born as an agent."
The headline is not just the 35-hour runtime, but the fact that the model continued to optimize long after other frontier models had plateaued. It signals that the Chinese AI ecosystem is moving from a strategy of "winning on cost" to "winning on raw capability."
Loading...