Youtu-VL-4B-Instruct-ncnn

Youtu-VL-4B-Instruct ncnn runtime package

Runtime-only pnnx/ncnn assets for the official tencent/Youtu-VL-4B-Instruct checkpoint pinned at 8d30a0e49662a1d628a472b12df264dbcd768753.

The package contains the SigLIP2 vision encoder, projector/merger, YoutuLLM embedding, 40-layer decoder with KV cache, final norm, LM head, and tokenizer. Inference uses the C++ youtu_vl_main executable and does not require Python, PyTorch, Transformers, or the original safetensors.

Example after building the model/youtu-vl branch:

youtu_vl_main \
  --root /path/to/this/package \
  --precision fp16 \
  --image testdata/coco_cats.jpg \
  --prompt "Describe this image in one sentence." \
  --full-decoder-ncnn \
  --max-new-tokens 3

The pinned regression fixture produces token IDs 10392, 5240, 1896 and the strict final text Two tabby in both the PyTorch baseline and ncnn runtime. See manifest.json for per-file sizes and SHA-256 digests.