Overview
- On Friday, August 14, Alibaba made both Qwen3.8 weight sets public with the 27B dense model released under Apache 2.0 and the 2.4 trillion-parameter Qwen3.8-Max published under a proprietary qwen3.8-max licence.
- The qwen3.8-max licence requires companies running 'model-as-a-service' or 'AI work assistant' businesses with over US$50 million in 12-month revenue to obtain a separate commercial licence while allowing internal use without that licence.
- Alibaba set per-token API prices of US$2 per million input tokens and US$6 per million output tokens for international access and 12/36 yuan per million tokens for China to steer high-volume commercial users toward paid services.
- Vendor reports claim the Max model can run at more than 4,000 tokens per second on Nvidia GB300 NVL72 hardware and scored 86.1 on an OSWorld-Verified benchmark, but those throughput and autonomy demos remain vendor-supplied and lack independent replication.
- Community groups rapidly produced conversion and quantization packs that make the 27B model practical on 12–32GB GPUs and Apple Silicon, which will broaden local experimentation and raise new questions about verification, governance, and cloud vs on‑premise economics.