Alibaba released Qwen3.8-Max on August 3: a 2.4-trillion-parameter MoE with a 1M-token context window, priced at $2/$6 per million tokens. The claim that got everyone's attention: 86.1 on OSWorld-Verified, ahead of GPT-5.6 Sol Max (83.2) and Claude Fable 5 (85.0) at agentic computer use. Open weights are promised "next week," alongside a 27B sibling.
I want this to be true. Frontier-class open weights would be the best thing to happen to indie builders since Llama. But I've been burned by launch-day benchmark euphoria before, so here's the checklist I actually run before moving any workload — using Qwen3.8-Max as the worked example.
1. Read the license before the model card
"Open weights" is doing a lot of work in that headline, because as I write this, the license hasn't been disclosed. Apache 2.0 and "custom license with a commercial-use clause that names your revenue threshold" are both routinely called open. Until there's a license file, there is no release — there's a press release.
If the weights ship under something restrictive, the 2.4T headline model matters less than whether the 27B sibling gets the permissive license. Which brings me to:
