We have published the well-known SmolVLM2-VL models for the RK3588 on Hugging Face.
There are three flavours: 2.2B, 500M, and 256M.
As you can see, they are pretty fast.
For the advanced users:
Porting the vision module of SmolVLM2 to rknn requires modifications to the intermediate ONNX file. Specifically, the indices for the Gather operation must be converted from a floating-point (FP16) data type to an integer. Please refer to the following image for an example.
Thank you for sharing your model — great work! It’s very impressive.
Would you mind if we use your model and GitHub repository as an example to create official documentation and publish it on docs.radxa.com? Of course, we will clearly credit you and link back to your repository.
