Qwen3.8-2.4T-A95B

Qwen3.8-2.4T-A95B is the checkpoint Alibaba uploaded to Hugging Face on 9 August 2026 — the model that powers the hosted Qwen3.8-Max, released for anyone to download. It is the first Max-class Qwen ever published this way and, at 2.446 trillion parameters, the largest set of downloadable model weights released by any laboratory to date. The architecture is a sparse mixture of experts: 92 layers holding 512 experts, of which ten routed experts plus one shared expert fire on each token, so roughly 95 billion parameters do the work at any moment. Three quarters of the layers use Gated DeltaNet linear attention and every fourth layer uses gated full attention, which is what keeps a 262,144-token context affordable to serve; Alibaba states the window can be stretched to 1,010,000 tokens. The practical catch is memory. At bfloat16 the weights alone occupy roughly 4.9 terabytes, and about 2.4 TB in the FP8 build published alongside — that is dozens of 80 GB accelerators before a single token of context is stored. The download figures show it: in its first month the 2.4T release and its FP8 twin were fetched about 44,000 times between them, while the small Qwen3.8-27B from the same family passed 3.4 million. The published model is also narrower than the rented one. Alibaba's own card states that hosted Qwen3.8-Max adds image input, a non-thinking mode, a million-token window by default and built-in tools; the download is text-in, text-out. Licensing is source-available rather than open: the bespoke Qwen3.8-Max License permits use, modification, hosting and resale, but a product above 100 million monthly users or 20 million US dollars in monthly revenue must display the model's name, and any model-as-a-service or AI work-assistant business earning more than 50 million dollars over twelve months must negotiate a separate licence first. The weights sat on Hugging Face for three days before that licence file was added on 12 August.

▸Videos

No videos yet.

▸Reviews