Alibaba unveils Qwen 3.7 Max, a fast, long‑running AI agent

Alibaba introduced its newest flagship model, Qwen 3.7 Max, a proprietary AI designed for long‑horizon, multi‑step tasks such as coding projects and enterprise automation. The model can execute more than 1,000 tool calls in a single run and runs about ten times faster than its predecessor, according to Alibaba’s internal testing [1].

Key takeaways

• Qwen 3.7 Max achieved ~35 hours of continuous autonomous execution, completing 1,158 tool calls in a single session [2].
• Benchmark scores include 92.4 on GPQA Diamond, 80.4 on SWE‑Verified, and 91.6 on LiveCodeBench, positioning it alongside top global models [1].
• The model supports OpenAI‑compatible APIs, Anthropic’s Claude Code, and OpenClaw, and offers a 1‑million‑token context window with a 64K output limit [2].
• Pricing on Alibaba Cloud Model Studio is $2.50 per million input tokens and $7.50 per million output tokens, plus $10 per 1,000 web‑search calls [2].
• The model is only accessible via Chinese endpoints, which may limit adoption by firms subject to data‑sovereignty regulations [2].

Long‑horizon reasoning and performance gains

Qwen 3.7 Max first appeared as a preview on the LM Arena leaderboard on May 14 and was formally announced at the Alibaba Cloud Summit in Hangzhou between May 19 and May 21 [1]. Alibaba credits iterative kernel optimizations for a ten‑fold speed improvement over earlier Qwen versions, enabling the model to sustain performance across hundreds or thousands of sequential actions—a capability it demonstrates by autonomously executing over 1,000 tool calls in internal tests [1].

The model’s endurance was highlighted in a demonstration where it ran for 35 hours on an isolated server with a T‑Head ZW‑M890 PPU, a hardware platform it had never seen during training. During this run it performed 1,158 distinct tool calls, evaluated 432 kernels, and achieved a 10.0× geometric‑mean speedup on an attention‑kernel optimization task [2]. Benchmark results show strong reasoning and coding abilities: 92.4 on GPQA Diamond (graduate‑level reasoning), 80.4 on SWE‑Verified (software‑engineering), and 91.6 on LiveCodeBench (coding on unseen problems) [1].

Compatibility, pricing, and market positioning

Qwen 3.7 Max is built for broad compatibility, supporting OpenAI‑style API formats and agent frameworks such as Claude Code and OpenClaw, and it natively handles the Anthropic API protocol [2]. Its massive 1‑million‑token context window and 64K output limit give developers ample space for large codebases and lengthy technical documents [2].

Access is provided through Alibaba Cloud Model Studio, with token pricing of $2.50 per million input tokens and $7.50 per million output tokens, plus a $10 fee per 1,000 integrated web‑search calls; code‑interpreter tools are free for a limited period [2]. This places Qwen 3.7 Max at roughly half the cost of comparable Western models such as OpenAI’s GPT‑5.4 or Anthropic’s Claude Opus 4.7, though its availability only from Chinese endpoints may deter enterprises bound by strict data‑locality rules [2].