Qwen posts FP8 build of its 2.4T-parameter flagship, Qwen3.8-2.4T-A95B-FP8

Qwen has published Qwen3.8-2.4T-A95B-FP8 to its official Hugging Face repository, tagging it as a quantized build of Qwen3.8-2.4T-A95B. The listing shows a text-generation pipeline, the transformers library, safetensors weights and a license marked “other”.

Registry metadata recorded 9,297 downloads and 261 likes at the time of collection, alongside tags for text-generation, conversational use and the qwen3_5_moe_text architecture. Those numbers are still modest for a flagship model, but they show the weights are already being pulled and tested rather than merely announced.

The name carries the specification. 2.4T is the total parameter count, A95B means roughly 95 billion parameters are active per token — the signature of a mixture-of-experts design — and FP8 describes 8-bit floating-point weights. This is not a new architecture but an official low-precision edition of the flagship base model.

Shipping FP8 is itself a statement about deployment. For teams running their own inference, the practical question is how much model fits on the hardware they already own and how many accelerators a given traffic load consumes, which matters more than a leaderboard number; an official release at least moves that trade-off into the vendor's own validation process.

It also continues this year's pattern of trillion-parameter open weights. The contest between open and closed models is shifting from whether a frontier-scale model exists in public at all to whether ordinary teams can afford to serve one.

One caveat sits in the license field, which reads “other” rather than a standard permissive license. Commercial users should read the repository terms for usage, redistribution and attribution requirements instead of assuming open weights mean unrestricted use.

Three things are worth watching: whether Qwen publishes accuracy comparisons against the unquantized base, how quickly inference frameworks and cloud providers offer turnkey support, and what real-world throughput and cost look like at this size.