Qwen3.7-Max runs autonomously for 35 hours to optimize code on its own custom chip

Qwen3.7-Max is Alibaba's proprietary push at agent foundations — explicitly designed for long-horizon reasoning rather than chat. The headline demo: 35 hours of unattended autonomous kernel optimization for code targeting Alibaba's own T-Head ZW-M890 PPU custom chip. That's a meaningful test of self-correction and tool use over multi-day horizons, comparable to claims OpenAI and Anthropic have made about background coding sessions.

The model is Alibaba Cloud Model Studio API-only (no open-weights release for Max), ships with a 1M-token context window, and targets agent frameworks — coding agents, office-task automation, and cross-framework consistency. The vertical-integration story is the strategic point: Alibaba is using its own model to optimize its own silicon, a flywheel that mirrors what xAI is doing with Colossus and Anthropic is starting with the NVIDIA Vera Rubin commit.

This lands the same week DeepSeek made its 75% V4-Pro price cut permanent and Alibaba and peers projected $84B in AI capex by 2027 — the Chinese frontier story is now about silicon plus model co-optimization, not just price.

Skeptics note that 35-hour autonomous runs are demo-grade until reproduced by external benchmarks. Watch next: Qwen3.7-Max on SWE-bench Verified, an open-weights Qwen3.7 base release, and how Alibaba prices Model Studio access vs. DeepSeek V4-Pro.