Alibaba released Qwen3.8-Max on Monday, and the company says it will publish the weights next week, making it the first Max-class Qwen model released openly. The model runs 2.4 trillion parameters with a context window of one million tokens, roughly 750,000 words per query. A second checkpoint, Qwen3.8-27B, is also going open-weights, sized for ordinary on-premise GPU hardware. Qwen3.8-Max builds on the Qwen3.5 architecture, using a mixture-of-experts design that activates 95 billion parameters per query out of the 2.4 trillion total. It takes text, images, video and documents as input, handles financial reports and PDFs past 200 pages, and can process videos longer than 100 hours. Pricing is listed at $2 per million input tokens, $6 per million output, and $0.25 per million for implicit caching, a flat rate across the entire million-token window. Alibaba published a full comparison table of benchmarks from its own runs. The generational jump is clearest in Terminal Bench moving 74.5 to 86.6, FrontierSWE 40.7 to 73.5, DeepSWE 21.6 to 56.6, and JobBench 31.3 to 53.4, while GPQA Diamond barely moved. It leads PaperBench at 93.0 against GPT-