minimind-v

Why we included this project

MiniMind-V is a rare thing in the open source VLM space: a full, minimal training pipeline you can actually run yourself rather than just a pile of pretrained weights. The project walks every step of building a small vision-language model, from a frozen SigLIP vision encoder that turns images into tokens, through the projector that aligns those features with a language model, to the Pretrain and SFT stages. A single NVIDIA RTX 3090 is enough to finish the SFT stage of the 65M model in about two hours, and the repo also ships several ready-made checkpoints down to 65M if you would rather load one and start experimenting. Because the code is deliberately stripped down and each stage is documented, it works as both a reference implementation for researchers and a hands-on starting point for engineers who want to reproduce a compact multimodal model on their own hardware. The cost and resource notes are a practical help for teams trying to figure out whether building small vision-language systems is worth it.