Qwen3.8-2.4T-A95B
This text-only mixture-of-experts checkpoint has 2.4 trillion total parameters and 95 billion active parameters. Its native context is 262,144 tokens.
Technical details
The architecture alternates gated delta networks with gated attention. The public weights do not add the hosted model's vision interface.Qwen3.8-2.4T-A95B model card
Practical use
Practical guidance. Estimate memory from the full stored parameter set, not the number of active experts. Route any commercial licence review to the actual intended service and revenue terms; the model's download access does not settle those conditions.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | Qwen/Qwen3.8-2.4T-A95BQwen/Qwen3.8-2.4T-A95B-FP8 |
| Architecture | 92 layers; 512 experts with ten routed experts and one shared expert.Qwen3.8-2.4T-A95B model card |
| Context extension | The model card documents extension to 1,010,000 tokens.Qwen3.8-2.4T-A95B model card |
| Licence | Custom Qwen3.8 Max terms. Qualifying MaaS or AI Work Assistant businesses above $50 million aggregate annual revenue need a separate commercial licence.Qwen3.8 Max licence |
Related models
- Qwen3.8-27B
Alibaba · Qwen3.8
Dense 27-billion-parameter vision-language model with weights released under Apache 2.0.
Launch date not confirmedDownloadable - Qwen3.8 Max
Alibaba · Qwen3.8
Hosted multimodal model with text, image and video input and a one-million-token context.
Launch date not confirmedAPI