Alibaba's Qwen3.7-Max Runs Autonomously for 35 Hours — A New Era for AI Agents

What did Alibaba's Qwen3.7-Max achieve?

Alibaba's Qwen Team released Qwen3.7-Max, a model that completed 35 straight hours of autonomous engineering work without human intervention. During this marathon session, it executed 1,158 distinct tool calls, performed 432 kernel evaluations, and achieved a 10x geometric mean speedup on an unfamiliar hardware architecture — the T-Head ZW-M890 PPU.

Why is 35-hour autonomy significant?

Most AI models degrade over extended sessions. They forget instructions, hallucinate variables, or get stuck in logical loops. Qwen3.7-Max was specifically designed as a "versatile agent foundation" for long-horizon reasoning. Maintaining coherent, productive work across 35 hours represents a fundamental shift from chat-based AI to true autonomous agents.

The proprietary shift

Unlike previous Qwen releases, Qwen3.7-Max is proprietary — not open source. This follows a broader industry trend where the most capable models are increasingly locked behind paid APIs. While this limits community access, it makes commercial sense for Alibaba given the enormous training costs involved.

Competition and compliance concerns

Qwen3.7-Max adds competition for American AI labs, which generally benefits consumers. However, since it's only accessible from Chinese-based endpoints, American and European enterprises may face data sovereignty and compliance barriers when considering adoption for government contracts.

What this means for the agent era

The AI industry has entered what many call the "agent era" — where models plan, execute, and course-correct complex tasks over days rather than seconds. Qwen3.7-Max's 35-hour run proves this is no longer theoretical. Expect more models to optimize for sustained autonomous operation.

FAQ

Q: Is Qwen3.7-Max open source? A: No. Unlike previous Qwen releases, Qwen3.7-Max is proprietary and available only through Alibaba's paid APIs.

Q: What task did it complete in 35 hours? A: It optimized an attention kernel on unfamiliar T-Head ZW-M890 PPU hardware, executing 1,158 tool calls and achieving a 10x speedup.

Q: Can Western enterprises use it? A: It's technically accessible but hosted on Chinese endpoints, which may raise data sovereignty and compliance concerns for regulated industries.

Stay ahead of the AI curve. Follow @AiForSuccess for daily insights.