Alibaba's Qwen team announced Qwen3.8-Max on August 3, following a preview shown in July at the World AI Conference in Shanghai. It's a sparse mixture-of-experts model with roughly 2.4 trillion total parameters but only about 95 billion active per token, which is the architectural trick that keeps inference costs manageable despite the model's huge total size. The context window tops out near 984K tokens, comfortably in "put a whole codebase in the prompt" territory, and it accepts multimodal input alongside text. For developers, the more interesting part is the built-in tool set: Qwen ships code_interpreter, web_search, web_extractor, and two image-search tools directly through its API, plus function calling, structured outputs, batch processing and fine-tuning support - meaning teams building agents don't have to wire up as much scaffolding themselves. On Terminal-Bench 2.1, a benchmark that scores how well a model can actually operate a terminal to complete real tasks, Qwen3.8-Max posted 86.6, trailing GPT-5.6 Sol's 88.8 but landing in the same tier as the current frontier coding models. Pricing is set at $2 per million input tokens and $6 per million output tokens, with cached input priced at $0.25 - competitive against Western frontier labs but pricier than some open alternatives. Alibaba says it will release open weights for both Qwen3.8-Max and a smaller Qwen3.8-27B variant within about a week of the announcement, though as of this writing neither is up on Hugging Face and no license has been specified. If the weights land as promised, this becomes one of the largest openly-released MoE models to date, which matters for teams that want to self-host a frontier-tier model rather than depend on a hosted API.