Dragonwing IQ-9075 (QCS9075) 1×HTP
99.8tok/s
TPS / Decode speed
76.1 ms
TTFT
883tok/s
Prefill speed
SmolVLM2-500M-Video-Instruct
Compact vision-language model that captions images and answers questions about them on-device, with text, image and video-frame input.
SmolVLM2-500M-Video-Instruct is a 500M-parameter instruction-tuned vision-language model. A SigLIP-style vision transformer turns each 512px tile into 1024 patches, which a pixel-shuffle and a linear connector compress to 64 tokens before they are spliced into the text sequence; a 32-layer grouped-query-attention decoder then generates the answer. Tokenization and image loading stay on the host; the compute backend depends on the hardware package (see hexagon/, cuda/, mlx/, amd/).
Model performance
SmolVLM2-500M-Video-Instruct on Hexagon HTP via refft-hexagon cli --model (single-file RFTB).
Measured · QCS9075
Install and run
Install
curl -fsSL https://raw.githubusercontent.com/refinefuture-ai/refft.cpp/main/refft-hexagon/install.sh | shRun
~/.local/share/refft-hexagon/bin/refft-hexagon cli --model ./SmolVLM2-500M-Video-Instruct-w4.refft --backend hexagon --prompt "Who are you?" --max_new_tokens 128Tags
llmvlmgenerative-aimultimodalquantized