Has anyone implemented a qwen-vl multimodal demo using expo?

[bbhxwl]

Has anyone implemented a qwen-vl multimodal demo using expo?

[a-ghorbani]

not Expo (RN CLI), but we are able to run Qwen3-VL-2B on a phone.

not entirely sure what issue you're running into, but in our case the problem was related to context window limits.

A couple of things to keep in mind:

Qwen-VL models use dynamic image sizing rather than a fixed size.

The number of output tokens scales based on image dimensions:

tokens = (H / effective_patch) × (W / effective_patch)

For Qwen3-VL, the effective patch size is patch_size × merge_size = 16 × 2 = 32 pixels (see preprocessor_config.json and clip.cpp#L3104).

for example for my phone's camera the image dims are 3024 × 4032, which means:

tokens = (3024 / 32) × (4032 / 32) ≈ 11,844 tokens

To manage this, llama.cpp caps output tokens at 4096 by default for Qwen-VL models (source). It calculates the maximum allowed pixels from this token limit, then scales the image down (calc_size_preserved_ratio).

So

• High-res images (common from phone cameras) can easily go above 4096 tokens (before the cap is applied)
• so, you need either:
  1. downscale the image before passing it to the model - use the formula above to calculate the n_tokens it nees
  2. or explicitly set a lower max token count and let it llama.cpp/mtmd to scale it: PR #275