Alibaba Launches Qwen3.8-Max, AI Model with 2.4 Trillion Parameters

Alibaba unveiled Qwen3.8-Max, an AI model built on a mixture-of-experts architecture containing 2.4 trillion total parameters. The model accepts text, image, and video inputs with support for a context window up to 1 million tokens. The hosted API is OpenAI-compatible and available immediately, while open weights will ship next week alongside the Qwen3.8-27B variant optimized for conventional datacenter hardware. In benchmarks, Qwen3.8-Max scored 86.6 on Terminal-Bench 2.1, surpassing Claude Opus 4.8 and Fable 5 (84.6). The model excelled in multimodal tasks, achieving 92.1 on OmniDocBench and 86.1 on OSWorld-Verified, demonstrating significant gains in agentic reasoning capabilities. It supports function calling, structured outputs, batches, and fine-tuning. Pricing is $2.00 per million input tokens and $6.00 per million output tokens. Caching reduces read costs by up to 87.5%, making prefix reuse economically advantageous.