Model comparison
| Model | Release | Max length | System prompt | Pretrained tokens | Q-LoRA memory | 2048-token Int4 generation | Tool use |
|---|---|---|---|---|---|---|---|
| Qwen-1.8B | 23.11.30 | 32K | Yes | 2.2T | 5.8GB | 2.9GB | Yes |
| Qwen-7B | 23.08.03 | 32K | No | 2.4T | 11.5GB | 8.2GB | Yes |
| Qwen-14B | 23.09.25 | 8K | No | 3.0T | 18.7GB | 13.0GB | Yes |
| Qwen-72B | 23.11.30 | 32K | Yes | 3.0T | 61.4GB | 48.9GB | Yes |
Values are mirrored from the upstream README tables and should be treated as source-cited historical product data.
Representative checkpoint paths
Variant cards
Qwen-1.8B
The smallest family member still ships 32K context and system prompt support in the chat variant.
Open linkQwen-7B
The most practical open deployment target in the original line, with base, chat, Int4, and Int8 checkpoints.
Open linkQwen-14B
The 14B release pushed the original line deeper into coding and Chinese knowledge while retaining tool-use support.
Open linkQwen-72B
The flagship open release in the original repo, combining 32K context, stronger system prompts, and the top benchmark results.
Open linkModel distribution and ecosystem
Model hubs
Mirror the official model cards across both ModelScope and Hugging Face so download paths are visible in both English and Chinese contexts.
Open linkDashScope API
The upstream README points to DashScope when you need a managed API surface instead of local model serving.
Open linkQwen-Agent
The tool-use and code-interpreter sections connect directly to Qwen-Agent for evaluation and agent workflows.
Open linkqwen.cpp
The original README highlights qwen.cpp as a lighter runtime path for the historical model line.
Open link