Alibaba's Qwen3.7-Max runs 35 hours autonomously

Alibaba said on May 21 that its new Qwen3.7-Max model is designed for “the agent era,” with the company highlighting a 35-hour autonomous run and compatibility with external coding harnesses that use Anthropic’s API style. The release puts Alibaba more directly into the market for long-running software agents, where model providers are competing on endurance, tool use and integration with existing developer workflows. VentureBeat reported the launch on May 21 and said the model can plug into external harnesses such as Anthropic’s Claude Code. ### What exactly did Alibaba say the model did for 35 hours? Alibaba said Qwen3.7-Max completed a “35-hour, fully autonomous kernel optimization run” involving more than 1,000 tool calls. In the company’s post, the run was presented as evidence that the model can maintain coherent reasoning over long sequences of actions rather than short chat exchanges. (alibabacloud.com) Alibaba’s description ties that result to coding work, not a general demo. The company said Qwen3.7-Max is built to write and debug code, automate office workflows and sustain execution across “hundreds or thousands of steps.” ### Why does Anthropic API compatibility matter here? VentureBeat reported on May 21 that Qwen3.7-Max supports Anthropic’s API protocol in tests, allowing it to work with tool integrations and external harnesses built around that interface. (alibabacloud.com) The article said that includes systems such as Claude Code, which developers use to connect models to coding tools and execution loops. Alibaba’s own post said the model “generalizes across agent scaffolds” and performs consistently when deployed through Claude Code, OpenClaw, Qwen Code and other frameworks. That means Alibaba is presenting the model as usable inside toolchains that were not originally built around Qwen. ### Is this an Alibaba Cloud launch or a research preview? (venturebeat.com) Alibaba said Qwen3.7-Max is a proprietary model and that it is “available soon” through Alibaba Cloud Model Studio. The company’s announcement did not present open weights in the material reviewed here, instead framing the release as an API product for agent workloads. May 21 was also the date of Alibaba’s public product post on Alibaba Cloud Community, which described Qwen3.7-Max as the latest top-end model in its lineup for coding, productivity and multi-agent orchestration. (alibabacloud.com) ### What is Alibaba claiming beyond the headline demo? Alibaba said Qwen3.7-Max can function as a coding agent, an office assistant through MCP integrations and a model for multi-agent orchestration. (alibabacloud.com) The company also published benchmark notes covering Terminal-Bench 2.0, SWE-Bench series, NL2Repo, MCP-Mark and other evaluations, though the announcement centered more heavily on agent behavior and scaffold compatibility than on a single benchmark table. VentureBeat said the broader implication is model interchangeability inside developer toolchains. That framing comes from the publication, not Alibaba, and reflects how support for Anthropic-style protocols could lower friction for teams that want to swap the underlying model without rebuilding the surrounding harness. (alibabacloud.com) ### What should developers watch next? Alibaba said Qwen3.7-Max is coming to Alibaba Cloud Model Studio, which is the next concrete release point named in the company’s May 21 post. Developers looking to test the model’s long-run behavior or Claude Code-style compatibility will need to wait for that availability step and the accompanying API access terms. (alibabacloud.com) (venturebeat.com)