Editorial illustration for Alibaba's Qwen3.7-Max runs 35 hrs, self‑monitors reward‑hacking, supports Claude Code
Alibaba's Qwen3.7-Max runs 35 hrs, self‑monitors...
Alibaba’s Qwen 3.7-Max, a large language model marketed as a “versatile agent foundation,” completed a 35-hour test on an unfamiliar processor last week. The model, working on an isolated server with a T-Head ZW-M890 PPU, was told to optimize an attention kernel. It made 1,158 tool calls and 432 kernel evaluations, achieving a 10.0× geometric-mean speedup.
Chinese competitor models GLM-5.1 and Moonshot’s Kimi K2.6 scored 7.3× and 5.0×, and sometimes stopped working altogether. Alibaba says the model was trained using “environment scaling” across a wide variety of agent scenarios, a process it credits for the result. The company also says Qwen 3.7-Max is built to work with external tools, including Anthropic’s Claude Code.
Over the course of 35 straight hours, Qwen3.7-Max operated entirely autonomously. It executed 1,158 distinct tool calls, performed 432 kernel evaluations, diagnosed compilation failures, and iteratively improved the code to achieve a 10.0x geometric mean speedup.
For engineers, the appeal of a model that can self-monitor for reward-hacking is fewer manual safeguards. The ability to work with tools like Claude Code also suggests flexibility for building hybrid systems. But the model’s 35-hour runtime, while a significant claim, was a single test in a controlled environment.
Other language models have been known to forget instructions or get stuck in repetitive loops over long sessions. The broader question is whether Qwen 3.7-Max can avoid those pitfalls consistently in real-world use. Alibaba has not yet published public benchmarks for independent verification.
Further Reading
- Qwen3.7: The Agent Frontier - Qwen Blog
- Qwen3.7-Max Benchmark Scores - Weights & Biases
- Qwen3.7 Max First Test – Hands-On With Alibaba's SMARTEST AI Model - YouTube
- Qwen3.7-Max: The Agent Frontier - Hacker News
Related Reading
Browse all Anthropic news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source
Related Topic Claude Mythos highlights EU AI safety gaps, says researcher Caroli
Policy & Regulation
From Archives Gemini 3.5 Flash Shows Fast Responses in Free Account Tests
From Archives Pricing Change Alters Complaint Language, Skews Classifier Accuracy
From Archives Claude skill helps data scientists spot 5‑6 PM weekday usage spikes in 2026
From Archives Deepseek launches Deepseek Code to compete with Claude Code and OpenAI's Codex
Common Questions Answered
What was Alibaba's Qwen 3.7-Max able to accomplish during its 35-hour test run?
Qwen 3.7-Max completed a 35-hour continuous test on an unfamiliar T-Head ZW-M890 PPU processor while optimizing an attention kernel. During this extended session, the model made 1,158 tool calls and 432 kernel evaluations, achieving a 10.0× geometric-mean speedup, which significantly outperformed competing Chinese models like GLM-5.1 (7.3×) and Moonshot's Kimi K2.6 (5.0×).
How does Qwen 3.7-Max's self-monitoring capability for reward-hacking benefit engineers?
The model's ability to self-monitor for reward-hacking reduces the need for engineers to implement manual safeguards and oversight mechanisms. This built-in monitoring feature allows for more autonomous operation while maintaining control over the model's behavior and decision-making processes.
What flexibility does Qwen 3.7-Max provide for system integration and development?
Qwen 3.7-Max supports integration with tools like Claude Code, which enables engineers to build hybrid systems combining multiple AI models and tools. This compatibility suggests the model can work alongside other specialized tools and models for more complex and flexible AI applications.
What are the limitations and concerns regarding Qwen 3.7-Max's 35-hour runtime performance?
While the 35-hour runtime is a significant achievement, it was conducted as a single test in a controlled environment, which may not represent real-world performance. Other language models have demonstrated issues with forgetting instructions or entering repetitive loops during extended sessions, raising questions about whether Qwen 3.7-Max can consistently avoid these pitfalls in varied conditions.
Further Reading
- Qwen3.7: The Agent Frontier — Qwen Blog
- Qwen3.7-Max Benchmark Scores — Weights & Biases
- Qwen3.7 Max First Test – Hands-On With Alibaba's SMARTEST AI Model — YouTube
- Qwen3.7-Max: The Agent Frontier — Hacker News
