Skip to content
Computer vision

TinyForge-Vision-222M

a compact multimodal model for visual question answering

2 min readupdated

TinyForge-Vision-222M is an ultra-lightweight 222.1M-parameter VLM (vision-language model that understands images) from developer Juanillaberia. It connects the google/siglip-base-patch16-224 image encoder (85.8M) and the HuggingFaceTB/SmolLM2-135M language model (134.5M) via a pixel-shuffle projector (compressing 2×2 patches into a single token). Built as a research and educational reproduction of nanoVLM v0.1 and trained on the Cauldron dataset, it answers short questions about images and describes their contents in English, though it trails much larger VLMs by a wide margin. It is a solid fit for developers and students studying multimodal architecture and training from scratch.

Key features

  • Compact footprint: just 222.1M parameters total (85.8M vision encoder, 1.8M projector, and 134.5M language model).
  • Image compression into 49 tokens of dimension 576, fed into the language model ahead of the text tokens (up to 79 text tokens).
  • Full unfreezing during training on 1.65M examples across 45 subsets of the Cauldron dataset at 224×224 resolution.
  • Strict prompt format: Question: {question} Answer: with no trailing space after the colon, matching the nanoVLM v0.1 codebase.

What you can do

  • Get short answers to simple visual questions or detailed scene descriptions in English.
  • Use the checkpoint and open repository as a hands-on reference for assembling and training a VLM from scratch.
  • Run experiments on prompt formatting and fine-tuning ultra-compact multimodal models.

Example

from tfv.models.vlm import VisionLanguageModel

model = VisionLanguageModel.from_pretrained("Juanillaberia/TinyForge-Vision-222M")
Read moreDetails
Metrics: 27.6% ± 1.1 on the MMStar benchmark (1,500 multiple-choice questions; compared to 35.3% for the published nanoVLM-222M — the gap comes down to the missing space after Answer: in the v0.1 code).
Training: 5 epochs (32,330 steps, batch size 256, bf16) took 250 minutes on a single NVIDIA A100 80 GB in Google Colab; the best checkpoint was selected by MMStar score (~1.9 epochs).
Limitations: English only; resizing images to 224×224 without preserving aspect ratio hurts small text reading, documents, charts, and object counting.
Inference: runs through the custom tinyforge-vision repository code rather than the standard transformers library.
Discuss Telegram 0 30

Will it run?

Runs in F32

Fits without quantisation.

  • F32weights 0.8 GB · ready files1.4 GB
  • BF16weights 0.4 GB1.0 GB
  • FP8 / INT8recommendedweights 0.2 GB0.7 GB
  • 4-bitweights 0.1 GB0.6 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · F32 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # Juanillaberia/TinyForge-Vision-222M
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download Juanillaberia/TinyForge-Vision-222M --local-dir ~/models/TinyForge-Vision-222M
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to TinyForge-Vision-222M
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of siglip-base-patch16-224 (quants, finetunes, LoRAs) will show up under the bell.

Quantised versions 9All on Hugging Face
Fine-tunes 12All on Hugging Face

About the model

Juanillaberia/TinyForge-Vision-222M
Author
Juanillaberia
Category
Computer vision
Released
Posted

On NeyroLearn

Hugging Face

300

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments