Qwen Releases Qwen3.8-2.4T-A95B MoE Model with FP8 Quantized Version

Qwen has published a new model, Qwen3.8-2.4T-A95B, on Hugging Face, marking the latest update to the open-source Qwen family.

The model card lists the pipeline as text-generation and shows it is built on the transformers library, with tags including qwen3_5_moe_text indicating a mixture-of-experts (MoE) text model.

Under Qwen's naming convention, 2.4T points to roughly 2.4 trillion total parameters and A95B to about 95 billion active parameters per inference, a classic sparse-activation MoE design.

At nearly the same time, the team uploaded an FP8 quantized variant, Qwen3.8-2.4T-A95B-FP8, whose model card lists the original release as its base model; quantized versions typically offer friendlier memory footprints and faster inference.

At the time of writing, the base model had recorded 978 downloads on Hugging Face while the FP8 version had around 3,851, with community interest climbing quickly.

The steady stream of large-parameter open-source MoE models lets developers run near-frontier text generation locally or in self-hosted setups while keeping inference costs in check through sparse activation.

Next to watch is whether the team publishes a technical report and benchmark results, and how quickly the model gains support in mainstream inference frameworks and deployment tooling.