Alibaba: Qwen3.7-Max Redefines Autonomous Chip Optimization

Alibaba's Qwen3.7-Max AI Model

Alibaba's Qwen team has recently launched an innovative artificial intelligence model, Qwen3.7-Max, which stands out for its ability to perform tasks autonomously. This model is specifically designed to handle complex software projects and is accessible only via an API, without a direct user interface. In practical tests, Qwen3.7-Max demonstrated its capability to optimize code entirely on its own, surpassing many competing models in terms of speed.

Qwen3.7-Max was designed to operate at the same level as leading AI labs in standard benchmarks. Developers have also utilized this model to independently detect undesirable behaviors and cheating attempts during its training process.

A 35-Hour Optimization Experience

In a landmark experiment, Qwen3.7-Max was tasked with optimizing a hardware-based attention kernel for the open-source inference software SGLang. The hardware used for this task was a cloud instance equipped with T-Head-ZW-M890 accelerators, an AI chip platform developed by Alibaba's semiconductor branch.

The Qwen team emphasized that the model had never been exposed to this chip architecture during its training. It began its work without measurement data, hardware documents, or example code, relying solely on the existing reference implementation written in Triton.

Over a continuous 35-hour period of autonomous work, Qwen3.7-Max executed 432 kernel tests and made a total of 1,158 tool calls. It compiled, measured, and revised the code iteratively, detecting compilation errors and identifying performance bottlenecks autonomously. Qwen researchers reported that the model achieved an average speedup of 10 times compared to the reference implementation.

Competing models yielded significantly lower results in the same configuration. For example, GLM 5.1 achieved a speedup of 7.3 times, Kimi K2.6 reached 5 times, DeepSeek V4 Pro managed 3.3 times, and the predecessor Qwen3.6-Plus barely progressed with a gain of 1.1 times. Models that dropped out early completed their sessions after five consecutive rounds without tool calls. On the standardized benchmark KernelBench L3, Qwen3.7-Max produced accelerated kernels 96% of the time, just behind Anthropic's Opus 4.6 at 98%.

Autonomous Training and Monitoring

Qwen3.7-Max relies on a training approach that the team initially implemented with Qwen3.5. Each training task is divided into three independent parts: the actual task, the tool environment, and the validator that checks the result. These elements can be freely combined, allowing the model to adopt effective strategies in various contexts.

The model was also used as a guardian during its own training. Over a period of more than 80 hours, Qwen3.7-Max monitored training sessions for software engineering tasks, performing over 10,000 checks. It looked for tricks that the training model might use to manipulate its rewards, such as obtaining correct answers directly from GitHub. In total, Qwen3.7-Max wrote 13 new detection rules and reported 1,618 cases of potential cheating.

Simulation Testing and Long-Term Planning

To assess its long-term planning capabilities, the team used YC-Bench, a benchmark that simulates the complete lifecycle of a startup over a year. The model must manage personnel through hundreds of decision rounds, review contracts, identify bad-faith clients, and maintain healthy profit margins in the face of rising labor costs.

Qwen3.7-Max generated a total revenue of $2.08 million and completed 237 tasks. Its predecessor, Qwen3.6-Plus, achieved $1.05 million, while Qwen3.5-Plus managed only $352,000.

In most benchmarks, Qwen3.7-Max is compared to Claude Opus 4.6 Max, Kimi K2.6 Thinking, GLM-5.1 Thinking, and DeepSeek V4 Pro Max. On SWE-Verified, the model scored 80.4, nearly on par with Opus 4.6 Max (80.8) and DeepSeek V4 Pro Max (80.6). In the math and science benchmarks GPQA Diamond (92.4), HMMT February 2026 (97.1), and Apex (44.5), Qwen3.7-Max outperformed the provider's comparison table.

Qwen3.7-Max is generally leading or tied with Claude Opus 4.6 Max, DeepSeek V4 Pro Max, GLM-5.1, Kimi K2.6, and its own predecessor Qwen3.6-Plus across twelve benchmarks. However, Claude Opus 4.6 remains the winner on NL2Repo, ClawEval, and CoWorkBench.

As the number of training environments increases, Qwen3.7-Max-Thinking climbs in the rankings across eight benchmarks, surpassing DeepSeek V4 Pro Max, GLM-5.1, and Kimi K2.6—but remaining just below Claude 4.6 Opus Max, according to the Qwen team.

Some of these evaluations are internal. QwenWebDev, QwenClawBench, CoWorkBench, and QwenWorldBench all come from the Qwen team itself. Each result here is self-reported. A deeper examination of scaling dynamics and methodology will be presented in an upcoming technical report.

Beyond typical use cases, the team also demonstrated Qwen3.7-Max controlling a four-legged robot. Using its own robotics framework and an associated navigation model, the language model guides the robot through physical spaces.