Following up on a recent post — Alibaba has fully released Qwen 3.7-Max 🐉

I've been using GLM-5.1 as my default chatbot for everyday tasks for quite a while. But apparently, I'm switching to the new Qwen.

Why Qwen 3.7-Max is currently (will it last at least a week?) sending GLM-5.1 (and many other models) to the bench:

🟢 Total dominance in reasoning. On the hardcore GPQA Diamond, Qwen scores 92.4 vs 86.2 for GLM-5.1. In tests for tool use (MCP-Mark) and complex engineering (SWE-Pro), the gap is even more noticeable.

🟢 Survivability in long-horizon tasks
Most models fall into hallucinations after a dozen tool calls. Alibaba tested Qwen3.7-Max on autonomous kernel optimization for unfamiliar hardware. The model poked at the code for 35 hours straight, made 1158 tool calls (profiling, compilation, debugging) and achieved a 10x speedup.
For comparison: on the same task, GLM-5.1 collapsed at 7.3x, and DeepSeek V4 Pro gave up entirely at 3.3x, simply stopping calling tools out of desperation.

🟢 Protection against cheating. Remember reports about how a model reverse-engineered the cache to pass tests? Qwen has built-in self-monitoring for SWE tasks. The model itself catches attempts at reward hacking (e.g., when an agent tries to sneak onto GitHub for a ready answer), writes heuristics for itself, and slaps its own hands.

Available in the web interface Qwen Studio and via API in Alibaba Cloud 👈🏻