Qianfan-VL-3B
Qianfan-VL-3B combines an InternViT image encoder with Qwen2.5-3B. It uses dynamic image patches for inputs up to 4K and has a documented 32K context.
Model design and access
The 3B release does not use the chain-of-thought feature described for the larger variants. The included Qwen terms restrict the covered Qwen materials to research and evaluation unless separate commercial permission is obtained.Qianfan-VL-3B official model card
Practical use
Practical guidance. Use the smaller checkpoint for constrained image-understanding experiments. Check the combined license before a commercial deployment. Test fine text and small visual objects at the exact image resolution used by your application.Qianfan-VL-3B official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Context window | 32,768 tokens Documented checkpoint context window; includes input and generated text. Qianfan-VL-3B official model card |
| License | MIT terms for Baidu contributions plus Qwen Research License for Qwen materialsQianfan-VL-3B repository license |
| Language-model ancestry | Qwen2.5-3BQianfan-VL-3B official model card |
Related models
- Qianfan-VL-70B
Baidu · Qianfan-VL
Seventy-billion-parameter Qianfan vision-language model built from Llama 3.1 and an InternViT image encoder.
Launch date not confirmedDownloadable - Qianfan-VL-8B
Baidu · Qianfan-VL
Eight-billion-parameter vision-language model combining Llama 3.1 and InternViT with a 32K context.
Launch date not confirmedDownloadable