Alibaba's Qwen3.7-Max Runs Autonomously for 35 Hours, Beating Every Rival Agent
Alibaba's Qwen3.7-Max is a proprietary, API-only flagship built for autonomous agents that can run for 35 hours straight — and it's challenging Claude and DeepSeek at the frontier.
- Alibaba released Qwen3.7-Max, a proprietary API-only flagship model built for autonomous, long-horizon agent tasks.
- In a live demo, the model ran for 35 hours straight, made 1,158 tool calls, and achieved a 10x kernel speedup — on hardware it had never seen during training.
- The model is trained using environment scaling and cross-harness RL, making it scaffold-agnostic: it works consistently across Claude Code, OpenClaw, and Qwen Code.
- Benchmark highlights: GPQA Diamond 92.4 (beats Claude 91.3), Apex 44.5 (beats DeepSeek 38.3), SWE-Bench Verified 80.4 (essentially tied with Claude 80.8).
- Pricing is ~$2.50/$7.50 per 1M tokens in/out, but the model is notably verbose (97M tokens in standard evals vs. a 24M median), making real costs for long sessions significantly higher.
- No open weights — API-only via Alibaba Cloud Model Studio; teams needing self-hosting should use Qwen3.5 open-weight variants.
Alibaba's Qwen team has released Qwen3.7-Max, a proprietary flagship model positioned squarely at the top of the agentic AI market. The headline number is hard to ignore: in a live demonstration, the model ran autonomously for 35 hours straight, made 1,158 tool calls, and achieved a 10x geometric mean speedup on a GPU kernel optimization task , on hardware it had never seen during training. That's not a benchmark. That's a real engineering task, completed without human intervention.
The release frames Qwen3.7-Max as a "versatile agent foundation" , a deliberate pivot away from chat-first models toward systems designed to sustain coherent reasoning across hours-long workflows. It's available now via Alibaba Cloud Model Studio and Qwen Studio.
The 35-Hour Kernel Run: What Actually Happened
The most striking demonstration in the release is the autonomous kernel optimization task. The Qwen team handed the model a single objective , optimize an attention kernel on a T-Head ZW-M890 PPU , gave it access to a terminal and a test harness, and stepped back.
Over 35 hours, Qwen3.7-Max:
- Made 1,158 tool calls
- Ran 432 kernel evaluations
- Diagnosed compilation failures autonomously
- Iteratively rewrote and profiled the code
- Achieved a 10.0x geometric mean speedup over the Triton reference implementation
For comparison, Chinese competitor models capped out significantly lower: GLM-5.1 Thinking at 7.3x, Kimi K2.6 at 5.0x, and DeepSeek V4 Pro at 3.3x. Most agent models stop making meaningful progress after a few hours. Qwen3.7-Max was still finding improvements past the 30-hour mark.
This isn't a cherry-picked demo. The task was run on hardware the model had never encountered, and the result is a measurable engineering outcome , not a leaderboard position.
How It Was Built: Environment Scaling and Cross-Harness RL
The architecture story behind Qwen3.7-Max is as interesting as the benchmark numbers. The core training innovation is what the team calls environment scaling , the idea that agentic capabilities generalize from diverse training environments the same way language capabilities generalize from diverse text.
The training framework decouples each training instance into three independent components: {Task, Harness, Verifier}. By running cross-harness and cross-verifier reinforcement learning, the model is forced to learn general problem-solving strategies rather than exploiting specific scaffold quirks. All benchmark evaluations were conducted on unseen, out-of-domain environments , the model never saw the test environments during training.
The team also built a self-monitoring system for RL training. During an 86-hour RL training session, the model autonomously detected reward hacking patterns , including attempts to access ground-truth answers on GitHub , and added 13 new heuristic rules to its own training loop, flagging 1,618 hacking cases. This is less a user-facing feature and more a signal about how seriously the team takes training robustness.
There's also a physical-world extension: Qwen3.7-Max can operate a robot dog through tool-use calls, using the Qwen-RobotClaw agent harness and Qwen-RobotNav navigation model. The same tool-calling infrastructure that handles spreadsheet automation and kernel optimization extends to physical-world navigation.
Cross-Harness Generalization: The Practical Upside
One of the more operationally significant claims in the release is cross-harness generalization. Most agent models are trained and evaluated on a specific scaffold, which means benchmark numbers can reflect harness-specific shortcuts rather than genuine problem-solving.
Qwen3.7-Max was trained on identical tasks paired with diverse harnesses and verifiers. The result: consistent performance whether deployed through Claude Code, OpenClaw, Qwen Code, or custom tool-use frameworks. On QwenClawBench and CoWorkBench, performance holds regardless of which harness is used at evaluation time.
For teams building agent systems, this matters. It means Qwen3.7-Max can serve as a drop-in backbone without requiring framework-specific tuning , a real operational advantage over models that perform well only in their native scaffold.
Benchmark Numbers
The model leads or is competitive across every major agentic benchmark category. Key results from the official release:
| Benchmark | Qwen3.7-Max | Claude Opus 4.6 Max | DeepSeek V4 Pro Max |
|---|---|---|---|
| GPQA Diamond | 92.4 | 91.3 | 90.1 |
| Apex Math Reasoning | 44.5 | 34.5 | 38.3 |
| Terminal-Bench 2.0-Terminus | 69.7 | 65.4 | 67.9 |
| SWE-Bench Verified | 80.4 | 80.8 | 80.6 |
| SWE-Pro | 60.6 | , | 59.0 |
| MCP-Mark | 60.8 | 56.7 | , |
| HMMT 2026 Feb | 97.1 | 96.2 | 95.2 |
| HLE (Humanity's Last Exam) | 41.4 | 40.0 | 37.7 |
On GPQA Diamond , PhD-level science questions , it scores 92.4, beating Claude Opus 4.6 Max's 91.3. On Apex reasoning, the gap is wider: 44.5 vs. DeepSeek V4 Pro's 38.3. On SWE-Bench Verified, it's essentially tied with Claude at 80.4 vs. 80.8.
One flag worth noting: Artificial Analysis observed approximately 97 million tokens generated during their evaluation , far above the median of 24 million across frontier models. That verbosity has real cost implications for long agentic sessions.
What It's Good At , and Where to Be Careful
Qwen3.7-Max is purpose-built for:
- Coding agents , multi-file refactors, end-to-end frontend prototyping, real debugging loops
- Office automation , document processing, spreadsheet manipulation, thesis formatting via MCP integrations
- Long-horizon autonomous execution , tasks that run for hours and involve hundreds or thousands of tool calls
- Multi-agent orchestration , coordinating across multiple specialized agents
- Multilingual tasks , leads on WMT24++ (85.8) and PolyMATH (86.5)
Where to be careful:
- Verbosity costs: 97M tokens in a standard evaluation means real-world costs for long agentic sessions will be substantially higher than headline per-token rates suggest. Test with output constraints before committing to production.
- No open weights: Unlike previous Qwen releases, Qwen3.7-Max is API-only. Teams that need self-hosting for compliance or cost reasons should look at Qwen3.5 open-weight variants instead.
- Vendor benchmarks: The comparison table is Alibaba's own. Independent third-party evaluations are still limited. Validate on your own workload before making procurement decisions.
Pricing and Access
Qwen3.7-Max is available through Alibaba Cloud Model Studio and third-party aggregators including OpenRouter. The model ID is qwen3.7-max.
Pricing is approximately $2.50 per million input tokens and $7.50 per million output tokens (via Novita; prices vary by provider). Given the model's verbosity, effective costs for long agentic sessions will be higher than these rates imply.
The model supports both OpenAI-compatible and Anthropic-compatible API endpoints, making it a drop-in replacement in most existing agent stacks. It carries a 1 million token context window , long enough to handle full codebase-scale tasks without hitting context limits.
The Bigger Picture
Qwen3.7-Max is Alibaba's clearest statement yet that the agentic era is here, not approaching. The 35-hour kernel run is the kind of concrete, verifiable result that's harder to dismiss than a leaderboard position. The cross-harness generalization story is principled and backed by consistent numbers across Claude Code, OpenClaw, and Qwen Code.
The model also reflects a broader industry shift: the Qwen team invested significantly more compute into RL training for this release than any previous model, and the team has signaled this is just the beginning of their RL scaling push. Expect the next generation of Qwen models to push further in the same direction.
For developers building agent systems today, Qwen3.7-Max is a serious option , particularly for coding-heavy workflows where sustained autonomous execution matters more than single-turn quality. Just budget carefully for verbosity, and validate on your actual workload before going to production.