Alibaba has pulled back the curtain on the latest Qwen series of LLMs with the introduction of the Qwen3.7-Max variant, built specifically to handle complex reasoning, multi-step workflows, and long-term coding tasks.
Instead of sputtering out during long operations, this new model can run continuously for up to 35 hours and manage more than 1,000 tool calls in a row without dropping in performance (allegedly). It plays nicely with major agent frameworks like OpenClaw, Hermes Agent, Claude Code, Qwen Paw, and Qoder, allowing it to jump straight into jobs ranging from quick frontend prototyping to sprawling, multi-file software engineering projects.
To feed these compute-hungry workloads, Alibaba Cloud built the Panjiu AL128 Supernode Server, crammed with 128 AI accelerators inside a single rack. The server uses an ICN Switch 1.0 networking chip to drive single-rack bandwidth up to the petabyte-per-second scale, which keeps data moving fast enough to handle thousands of concurrent requests without bottlenecking.
On the silicon side, Alibaba semiconductor branch T-Head cooked up the Zhenwu M890, an AI processor that serves up three times the performance of the older Zhenwu 810E. The hardware packs 144GB memory alongside 800GBps inter-chip bandwidth, and it can scale its math from high-precision FP32 down to ultra-low-precision FP4 to keep costs low and execution snappy.
That processor links up with the new ICN Switch 1.0 chip, a dedicated switching component that pushes 25.6 terabits per second of aggregate bandwidth to connect up to 64 accelerators without congestion. Everything ties together through a proprietary software stack dubbed T-Head SAIL, which has already found a home in the market with over 560,000 Zhenwu units shipped to 400 external clients across 20 industries.
On the operational side, Alibaba platform Bailian is rolling out an Agentic RL reinforcement learning mechanism that uses real-world execution feedback to sharpen model accuracy over time. The platform also gets built-in safety guardrails to ensure autonomous tools stay within pre-set operational boundaries while running enterprise tasks.
The Qwen3.7-Max model will be available soon for global developers and enterprises through the Model Studio platform.
