Qwen3.7: A New Horizon for Agents

Qwen3.7-Max is the latest proprietary model designed for the agentic era, a general-purpose agent foundation model with capabilities spanning coding, office workflow automation, and autonomous execution over thousands of steps.

The model possesses the following core capabilities:

• Coding Agent: Performs everything from frontend prototyping to complex multi-file engineering.
• Office and Productivity: Automates workflows through MCP(Model Context Protocol) integration and multi-agent orchestration.
• Long-horizon Autonomous Execution: Capable of performing long-duration tasks, such as a case where it completed kernel optimization by making over 1,000 tool calls across 35 hours.
• Versatility: Delivers consistent performance across various agent frameworks including Claude Code, OpenClaw, and Qwen Code.

According to benchmark results, Qwen3.7-Max recorded overwhelming performance on key metrics such as SWE-Pro, Terminal Bench 2.0, GPQA Diamond, and HLE, surpassing or matching competing models such as Opus-4.6 Max and DS-V4-Pro Max.

Technically, the model employs an Environment Scaling approach that expands the quality and diversity of agent training environments to enhance generalization ability. In addition, it is designed with a Rollout environment infrastructure that separates Task, Harness, and Verifier, preventing bias toward optimization for a specific framework and enabling the model to learn consistent problem-solving strategies across diverse environments.

In practice, when placed in a new hardware environment it had never encountered during training (T-Head ZW-M890 PPU), the model demonstrated remarkable autonomy by autonomously optimizing the Extend Attention kernel of SGLang through 1,158 tool calls over 35 hours.