Qwen3.8-Max carries 2.4 trillion parameters and a context of one million tokens, already available through the API. The real news is different: Alibaba says that in days it will release the weights freely - a top-class model anyone will be able to download.
- 2.4 trillion parameters total, but only 95 billion active per token: a sparse Mixture-of-Experts architecture with hybrid attention; context up to one million tokens.
- Available through the API in Alibaba Cloud Model Studio; the weights are promised for release next week.
- Per Alibaba's own data: fifth place in Text Arena, second in Vision Arena, fourth in Frontend Code Arena.
Two weeks after Moonshot released K3's weights, Alibaba raises the stakes with the biggest model it has ever built - and says everyone will be able to download this one too.
The numbers are impressive, but the commercial logic is more interesting. The 2.4 trillion is the nameplate, the 95 billion active is the bill - you pay speed and power like a mid-size model, while the size sits in the back like a library. That exact arithmetic is what made open Chinese models dangerous to closed pricing: catching-up quality at a fraction of the price.
The open part here is, for now, only a promise - the weights aren't out yet, and that's exactly what I'll watch next week. If they really release them, every closed price list in the industry will have to explain exactly what it's selling above the free version. Until then, we're counting down the week.