Summary
On August 3, 2026, Alibaba released Qwen3.8-Max, a mixture-of-experts model with 2.4 trillion total parameters (about 95B active per query) and a 1M-token context window, natively accepting text, image, and video input. It launched via API at $2 and $6 per million input/output tokens, with open weights for Qwen3.8-Max and a 27B variant promised the following week, and Alibaba published benchmarks it says rival Anthropic's Fable 5.
What changed
Alibaba released Qwen3.8-Max, a 2.4T-parameter MoE model (~95B active) with a 1M-token context and text/image/video input, priced at $2/$6 per million tokens via API, with open weights due the following week.
Why it matters
A frontier-scale multimodal model from Alibaba, priced well below Western frontier models and paired with a promised open-weight release, raises the ceiling for non-US options and gives gateways another high-context model to route to. The MoE design keeps active-parameter serving cost down, making trillion-scale capability more deployable.
Evidence excerpt
Qwen3.8-Max is a Mixture-of-Experts model with 2.4 trillion total parameters and 95 billion parameters active per query... natively multimodal, accepting text, images and video... a context window of up to 1 million tokens... pricing of $2.00 and $6.00 per million tokens, with open weights for Qwen3.8-Max and Qwen3.8-27B due the following week.