Alibaba pushed two new Qwen models to public preview this week: Qwen 3.7 Max Preview and Qwen 3.7 Plus Preview, both already live for testing on Arena and inside Qwen Chat. Max Preview ranks 13th overall on the Arena Text leaderboard and 7th on math reasoning. Plus Preview ranks 16th overall on the Vision leaderboard. An official Alibaba Cloud Summit announcement is expected this week.
What this enables for creators
If your work depends on the open-weights or hosted Qwen line, you can stress-test 3.7 against your real prompts right now. Open the Arena leaderboard, anonymous-pair against your current LLM, and see whether the new generation handles your specific workload before the official release. The same goes for the vision side, where Plus Preview competes directly with the small MiniCPM-V 4.6 and other multimodal contenders. Lock in a baseline score now and you have data ready the day Alibaba opens API access.
Why it matters
Top-15 Arena placement is the practical floor for a flagship model in 2026. The Max Preview being 13th on Text and Plus Preview being 16th on Vision puts Alibaba at the number 6 lab on Text and number 5 lab on Vision globally, second only to Western frontier labs. SCMP reports these are the highest-ranking Chinese AI models on the leaderboard to date. The pace of Qwen iteration also keeps pressure on the Chinese open-weights ecosystem that the DeepSeek V4 family and Moonshot funding round have been driving.
Key details
Both models are preview-only and hosted, not downloadable open weights yet, which is the standard Alibaba sequencing for a flagship cycle. Specific Arena ranks: Max Preview is 13th overall on Text, 7th in math reasoning, 9th in expert applications, 9th in software and IT, and 10th in programming. Plus Preview is 16th overall on Vision. Community discussion on Hacker News at 240+ points speculates the new versions are finetunes on the 3.5 architecture rather than a from-scratch retrain, which explains the rapid 3.5 to 3.7 cadence. An open-weights drop typically follows the hosted preview by a few weeks if past Qwen cycles hold.
What to do next
Run your hardest 10 prompts through Qwen 3.7 Max Preview on Arena right now and capture the responses. Do the same for your image-and-text prompts on Plus Preview. Save a copy because preview model behavior shifts between drops. When Alibaba opens API access at the Summit, you have a calibrated comparison ready and a hard answer on whether 3.7 belongs in your stack alongside your existing Qwen-based agent workflows.
