Qwen (Alibaba) models
A documented index of Qwen (Alibaba) model milestones, specifications and available pricing.
Model index.
| Model | Lab | Dated | Latest event | Weights | Input / output modalities | Context | Input / output | Verified | Source |
|---|---|---|---|---|---|---|---|---|---|
| Qwen3.8-Flash-Next | Qwen (Alibaba) | released | Open weights | Text, image and video input; text output | 262,144 | — | Official ↗ | ||
| Qwen3.8-2.4T-A95B | Qwen (Alibaba) | released | Open weights | Text input; text output; thinking always on, vision not supported | 262,144 | — | Official ↗ | ||
| Qwen3.8-27B | Qwen (Alibaba) | released | Open weights | Text, image and video input; text output | 262,144 | — | Official ↗ |
Dated = a documented vendor milestone, not necessarily the first public announcement. Prices are USD per million text tokens where shown. Context varies by provider and interface; where a model has locally available weights, context may describe the official model card. “—” means not verified; check conditions and sources before use.
Recent milestones.
- released
Qwen3.8-Flash-Next
Experimental open-weights preview of the architecture Qwen says will underpin Qwen4: Qwen Sparse Attention, Gated Residual and n-gram embeddings; 125B total with 6B activated, under the qwen-community-1.0 license.
- released
Qwen3.8-2.4T-A95B
First Qwen-Max-class model released as open weights (2.4T total, 95B activated MoE) under the qwen3.8-max license. Text-only with thinking always on; native 262,144-token context, extensible to 1,010,000.
- released
Qwen3.8-27B
Open-weights release under Apache-2.0; 27B dense vision-language model for agentic coding and long-horizon work. Native 262,144-token context, extensible to 1,000,000. Qwen says a hosted Qwen Cloud version is coming soon.