Alibaba's Qwen3.7‑Max runs 35 hours autonomously, supports Claude Code

Alibaba unveiled its flagship Qwen3.7‑Max model, claiming it can operate autonomously for about 35 hours while making more than 1,000 tool calls, and it is compatible with external agent frameworks such as Anthropic’s Claude Code [2].

Key takeaways

• Qwen3.7‑Max completed roughly 1,158 tool calls in a 35‑hour continuous run on Alibaba’s own Zhenwu M890 accelerator [2].
• The model executes over 1,000 tool calls in a single session in internal tests, running about ten times faster than its predecessor [1].
• It supports popular API formats and agent frameworks, including Claude Code and OpenClaw [1].
• Benchmark scores place Qwen3.7‑Max among frontier‑tier models, with a GPQA Diamond score of 92.4 and LiveCodeBench score of 91.6 [1].
• The demonstration ties together Alibaba’s “AI factory” stack: the model, a purpose‑built chip, and a rack‑scale server [2].

Demonstration of long‑horizon autonomy

At the Alibaba Cloud Summit in Hangzhou (May 20‑21), the company showcased a full‑stack AI workflow. Qwen3.7‑Max was tasked with writing and iteratively optimizing software for Alibaba’s proprietary Zhenwu M890 accelerator, a chip it had not previously documented. After the model generated the software stack, it ran the same model on the newly optimized chip for an uninterrupted 35‑hour period, issuing 1,158 tool calls and 432 kernel evaluations [2]. Alibaba attributes the speed gains to iterative kernel optimizations that deliver an approximate ten‑fold improvement over earlier versions [1].

Compatibility and benchmark performance

Beyond the endurance test, Qwen3.7‑Max is designed for broad interoperability. It accepts OpenAI‑style API calls and integrates with agent frameworks such as Anthropic’s Claude Code and OpenClaw, enabling developers to embed the model in existing toolchains [1]. On a suite of benchmarks, the model achieved a 92.4 score on GPQA Diamond (graduate‑level reasoning), 80.4 on SWE‑Verified (software engineering), and 91.6 on LiveCodeBench (coding on unseen problems) [1]. Independent rankings from Artificial Analysis placed it at the top of Chinese models on a composite Intelligence Index, though its factual recall accuracy dropped relative to its predecessor [2].

Why it matters

The 35‑hour autonomous run underscores Alibaba’s strategy of vertical integration: by controlling chip design, model training, and deployment platforms, the company can fine‑tune AI workloads for its own hardware—a capability that could reduce reliance on foreign processors. The model’s ability to handle thousands of sequential tool calls positions it for complex enterprise use cases such as automated software pipelines, multi‑system customer service, and large‑scale financial reporting [1]. While the performance claims are currently based on Alibaba’s internal measurements, external validation will be needed to confirm the model’s competitiveness against leading Western offerings. Alibaba plans to make Qwen3.7‑Max available through its Model Studio API, with broader developer access “coming soon” [2]. Continued benchmarking and third‑party testing will determine how the model’s long‑horizon autonomy translates into real‑world productivity gains.