Alibaba drops Qwen3.7-Max to power autonomous agents

The benchmark for frontier AI has shifted. Nobody cares if a model can write a catchy poem anymore; the real test is whether it can sit in a terminal, fix buggy code, call tools, and keep working without hallucinating off a cliff.

Alibaba introduced Qwen3.7-Max, a flagship proprietary model explicitly engineered for the agent era. Instead of optimizing for simple conversational Q&A, Alibaba designed this variant to tackle complex, multi-step autonomous execution.

A 35-hour autonomous marathon

What immediately catches attention is the model's endurance over long reasoning horizons. Most LLMs lose the plot after a handful of agentic loops. Alibaba claims Qwen3.7-Max pulled off a 35-hour, fully autonomous GPU kernel optimization run that executed over 1,000 tool calls.

On the coding front, the benchmark numbers place it firmly among top-tier frontier systems. It scored 80.4 on SWE-Verified, putting it virtually neck-and-neck with Opus-4.6 Max at 80.8 and DS-V4-Pro Max at 80.6. On Terminal-Bench 2.0, it notched 69.7, outperforming DS-V4-Pro Max at 67.9.

Beyond command-line engineering, the model integrates vision natively. It can parse dense financial charts and video streams alike, with Alibaba showcasing its use as the navigation brain for robotic hardware.

Playing nice with everyone else's scaffolding

Alibaba took a practical approach to distribution. Rather than forcing developers into a proprietary walled garden, Qwen3.7-Max was tuned for cross-scaffold generalization. It drops right into popular environments like Claude Code, OpenClaw, and Hermes, alongside deep Model Context Protocol (MCP) support.

There is one major catch: this model is not open-source. For now, Qwen3.7-Max is only available via Alibaba Cloud Model Studio and paid APIs.

Why it matters

We think this release signals a clear transition in how leading labs deploy their flagship systems. Raw conversational intelligence has plateaued, but multi-hour autonomous agency remains wide open. By baking long-horizon error recovery directly into the model weights, Alibaba reduces the fragile glue code developers previously had to build.

Embracing third-party scaffolds like Claude Code and MCP is an exceptionally smart move. It allows Alibaba to siphon developers away from competing frontier APIs without demanding they rewrite their existing automation pipelines.

However, keeping this Max variant closed feels like a calculated risk. The global momentum behind Qwen was largely powered by its open-weights ethos. For teams that bet on Alibaba specifically for local sovereignty and self-hosting, being pushed back to a cloud API is a bitter pill. Moreover, letting an agent burn through 1,000 sequential tool calls could quickly generate jaw-dropping API bills if proper guardrails are not in place.