---

SmolVLM2-256M-Video-Instruct-GGUF

Original Model

HuggingFaceTB/SmolVLM2-256M-Video-Instruct

Run with LlamaEdge

• LlamaEdge version: v0.21.0 and above

• Prompt template

  • Prompt type: smol-vision

  • Prompt string

    <|im_start|>
    User: {user_message_1}<image>
    Assistant: {assistant_message_1}
    User: {user_message_2}<image>
    Assistant:
    

• Context size: 2048

• Run as LlamaEdge service

  wasmedge --dir .:. --nn-preload default:GGML:AUTO:SmolVLM2-256M-Video-Instruct-Q5_K_M.gguf \
    llama-api-server.wasm \
    --prompt-template smol-vision \
    --llava-mmproj SmolVLM2-256M-Video-Instruct-mmproj-f16.gguf \
    --model-name SmolVLM2-256M-Video-Instruct \
    --ctx-size 2048
  

Quantized GGUF Models

Name	Quant method	Bits	Size	Use case
SmolVLM2-256M-Video-Instruct-Q2_K.gguf	Q2_K	2	104 MB	smallest, significant quality loss - not recommended for most purposes
SmolVLM2-256M-Video-Instruct-Q3_K_L.gguf	Q3_K_L	3	114 MB	small, substantial quality loss
SmolVLM2-256M-Video-Instruct-Q3_K_M.gguf	Q3_K_M	3	110 MB	very small, high quality loss
SmolVLM2-256M-Video-Instruct-Q3_K_S.gguf	Q3_K_S	3	104 MB	very small, high quality loss
SmolVLM2-256M-Video-Instruct-Q4_0.gguf	Q4_0	4	108 MB	legacy; small, very high quality loss - prefer using Q3_K_M
SmolVLM2-256M-Video-Instruct-Q4_K_M.gguf	Q4_K_M	4	125 MB	medium, balanced quality - recommended
SmolVLM2-256M-Video-Instruct-Q4_K_S.gguf	Q4_K_S	4	122 MB	small, greater quality loss
SmolVLM2-256M-Video-Instruct-Q5_0.gguf	Q5_0	5	125 MB	legacy; medium, balanced quality - prefer using Q4_K_M
SmolVLM2-256M-Video-Instruct-Q5_K_M.gguf	Q5_K_M	5	133 MB	large, very low quality loss - recommended
SmolVLM2-256M-Video-Instruct-Q5_K_S.gguf	Q5_K_S	5	131 MB	large, low quality loss - recommended
SmolVLM2-256M-Video-Instruct-Q6_K.gguf	Q6_K	6	169 MB	very large, extremely low quality loss
SmolVLM2-256M-Video-Instruct-Q8_0.gguf	Q8_0	8	175 MB	very large, extremely low quality loss - not recommended
SmolVLM2-256M-Video-Instruct-f16.gguf	f16	16	328 MB
SmolVLM2-256M-Video-Instruct-mmproj-f16.gguf	f16	16	190 MB

Quantized with llama.cpp b5501