Qianfan-VL-70B
Qianfan-VL-70B combines an InternViT image encoder with Llama 3.1. It uses dynamic image patches for inputs up to 4K and has a documented 32K context.
Model design and access
The 70B release supports chain-of-thought responses. Although its card metadata says MIT, the repository license also includes the Llama 3.1 Community License.Qianfan-VL-70B official model card
Practical use
Practical guidance. Reserve enough memory for the full language model and its visual tokens. A larger checkpoint may improve difficult reasoning, but it does not remove OCR or localization errors. Check conclusions against the source image when they affect downstream records.Qianfan-VL-70B 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-70B official model card |
| License | MIT terms for Baidu contributions plus Meta Llama 3.1 Community LicenseQianfan-VL-70B repository license |
| Language-model ancestry | Llama 3.1Qianfan-VL-70B official model card |
Related models
- Qianfan-VL-3B
Baidu · Qianfan-VL
Vision-language model combining Qwen2.5-3B with an InternViT encoder for images up to 4K.
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