SmolVLM 500M Instruct (MNN)

Hugging Face's small vision-language model SmolVLM 500M Instruct as an MNN bundle, for describing and answering questions about images on a phone.

Published by TokForge for on-device inference.

What it is Vision chat model (text + images)
Format MNN bundle, 4-bit weights (block 64), 8-bit vision encoder, BF16 embeddings
Runs on MNN engine (CPU); the TokForge app uses taobao-mnn's copy of this model instead
Phone class No catalog class; needs more free RAM than its 0.43 GB of weights
Download 431 MB (10 files)
Base model HuggingFaceTB/SmolVLM-500M-Instruct
License Apache-2.0

TokForge

Runs on-device in the TokForge app.

Use it in TokForge

  1. Install TokForge from Google Play, or join the iOS TestFlight.
  2. You do not need to download this repo. When you attach a photo to a chat with a model that cannot read images itself, TokForge describes the photo with SmolVLM 500M Instruct and the chat model answers from that description. The app fetches that helper by itself from taobao-mnn/SmolVLM-500M-Instruct-MNN, the MNN team's export of the same model, not from this repo.

Collections: TokForge Vision Models, TokForge Voice and Helper Models.

Files

File Size What it is
config.json 387 B MNN runtime settings
embeddings_bf16.bin 94.6 MB Token embeddings (bf16)
export_args.json 1.08 KB Export settings used for the conversion
llm.mnn 558 KB Model graph
llm.mnn.json 1.12 MB Model graph as JSON (reference copy)
llm.mnn.weight 227 MB Model weights
llm_config.json 1.12 KB Model settings and chat template
tokenizer.txt 563 KB Tokenizer
visual.mnn 228 KB Vision encoder graph
visual.mnn.weight 107 MB Vision encoder weights

About the model

SmolVLM 500M Instruct is a small vision-language model by Hugging Face (HuggingFaceTB): it reads an image and answers questions about it in English. In MNN it uses the Idefics3 layout (model_type: idefics3 in llm_config.json).

How it was made

Converted with MNN's llmexport straight from HuggingFaceTB/SmolVLM-500M-Instruct: language model weights 4-bit with block size 64, vision encoder 8-bit, embeddings BF16 in embeddings_bf16.bin (export_args.json records the settings).

License

Apache-2.0, as Hugging Face releases SmolVLM 500M Instruct.

Credits

  • SmolVLM-500M-Instruct by Hugging Face (HuggingFaceTB) (Hugging Face). Used under the Apache-2.0 license. Built on SmolLM2-360M-Instruct and Google's SigLIP (Apache-2.0).
  • Converted with llmexport from MNN by Alibaba (Apache-2.0).
Downloads last month
33
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for darkmaniac7/SmolVLM-500M-Instruct-MNN

Collections including darkmaniac7/SmolVLM-500M-Instruct-MNN