other image sizes qwen3-VL-2B, structured output mode

General Discussion
1

Hi Hailo team,

We’re running Qwen3-VL-2B-Instruct from the GenAI Model Zoo (v5.3.0) on a Raspberry Pi 5 with the AI HAT+ 2 (Hailo-10H), using HailoRT 5.3.0. It works well, and we’d like help with two things:

  1. Other image sizes. The model only accepts 512x288 input. We’d like to run it at higher resolutions, such as 768 px or 1280x720, but we don’t see other sizes on the download page, and from what we’ve found online, compiling our own VLMs isn’t currently supported. Is there a version compiled for other sizes, or a way for us to get one?

  2. Structured (JSON schema) output. Describing the format in the prompt isn’t reliable enough: the output often drifts from it. What we need is a structured output mode, where we pass a JSON schema and generation is restricted so the output always matches it. We don’t see support for this in the documentation. Is it something you plan to support, or have available in some capacity, similar to the JSON schema option in frameworks like llama.cpp?

Thanks for your help