Loading...
Loading...
Alibaba on May 20 launched Qwen3.7-Max, a next-generation flagship large language model tuned for agent-style, long-horizon workflows. The model demonstrated strong benchmark performance across coding, reasoning, multilingual tasks and general agent metrics, and reportedly completed a 35-hour autonomous task with over 1,000 tool calls. Running on a new chip platform and optimized for GPU kernel acceleration, Qwen3.7-Max delivers roughly 10x faster inference than its predecessor and shows cross-framework code-generalization. Alibaba will offer the model via API on its Cloud Bailian platform, signaling a push to commercialize advanced agent capabilities for enterprise automation, AI assistants and cloud services.
Alibaba's Qwen3.7-Max signals a step change in agent-oriented LLMs, emphasizing long-horizon autonomous workflows and enterprise integration. Tech teams should note implications for automation, cloud AI offerings, and tooling around multi-tool agent orchestration.
Dossier last updated: 2026-05-20 12:38:22
Alibaba chair Joe Tsai and CEO Wu Yongming told shareholders that Alibaba’s AI business has moved beyond early investments into a commercialization phase and is becoming a material growth driver tied to AI+cloud. The letter details investments across the stack: self‑developed Pingtouge AI chips in mass production, accelerated foundation model work (Qwen3.7‑Max and the Qianwen family), specialized models (HappyOyster, HappyHorse), and productized offerings including the consumer Qianwen App and enterprise Wukong platform. Alibaba says cloud intelligent external revenues accelerated to 40% growth last quarter with AI making up 30% of that, and vows continued capex in infrastructure, chips, models and MaaS to connect models to applications—positioning AI to boost e‑commerce, instant retail and broader ecosystem monetization.
Qwen3.7-Max: The Agent Frontier
Qwen3.7-Max: The Agent Frontier
Alibaba on May 20 unveiled its next-generation flagship Qwen3.7-Max large language model, which topped Chinese models in the Arena global blind benchmark. Qwen3.7-Max is redesigned for agent-style workflows, claiming sustained gains in programming and reasoning and the ability to autonomously complete a 35-hour complex agent task. Running on a new chip platform, the model reportedly performed over 1,000 tool calls and self-evolved a critical kernel, delivering roughly 10x faster inference than the prior version. The announcement signals Alibaba’s push to lead domestic AI model performance and agent capabilities, with implications for cloud services, AI assistants, and enterprise automation.
Alibaba's Qwen team today released Qwen3.7-Max, a flagship agent-oriented large model that will be offered via API on Alibaba Cloud's Bailian platform. The model targets broad agent capabilities—coding, office automation, multi-agent workflows and sustained long-horizon autonomous execution—and demonstrated robustness in a 35-hour, 1,000+ tool-call continuous task. Benchmarks shared by the company show leading or competitive results across programming (SWE suites, SciCode, Terminal Bench), general agent metrics (MCP-Mark, MCP-Atlas, Skillbench), GPU kernel acceleration, reasoning (GPQA Diamond, HLE, HMMT), and multilingual tasks (WMT24++, IFBench). Qwen3.7-Max also claims cross-framework generalization (Claude Code, OpenClaw, Qwen Code) and will be accessible via Alibaba Cloud API.