Alibaba's Qwen3.7-Max, released in late May, ranks as the #1国产模型 (top domestic model) on the Arena blind leaderboard. It's built specifically for the Agentic era.
In a striking real-world test, Qwen3.7-Max ran autonomously for 35 hours on a new, unseen hardware platform (T-Head Zhenwu M890 chip). It conducted 432 kernel evaluations and 1,158 tool calls, ultimately delivering a 10x speedup over the SGLang Triton reference implementation.
On SWE-Pro and Terminal Bench 2.0, it outperforms DeepSeek-v4-pro-Max and Claude-Opus4.6. On MCP-Atlas (agentic tool orchestration), Qwen3.7-Max hits 75.3, close to GLM-5.2's 77.0 and Opus 4.8's 77.8.
Qwen-AgentWorld: Released June 24, this is the first native "Language World Model" covering seven agentic domains (MCP, Search, Terminal, SWE, Web, OS, Android) in a single model. On AgentWorldBench, the 397B-A17B version surpasses GPT-5.4, Claude Opus 4.8, and Gemini 3.1 Pro in overall simulation quality.
Qwen3.7-Plus: A multimodal model ranking among the global top 5 on Vision Arena (#1 in China). It integrates "see, think, write, do, verify" into a unified agentic workflow.
Top comments (1)
niu