Most vision-language models arrive as opaque checkpoints measured in billions of parameters; this one ships as a training recipe you can run on the GPU already sitting in your desktop. The real product isn't the 65M model it produces — far too small to rival production VLMs — but the fully readable PyTorch pipeline that shows exactly how pixels turn into tokens a language model can reason over.
What Sets It Apart
- The whole loop — a frozen SigLIP2 vision encoder, a two-layer MLP projection, and a small MiniMind language backbone — trains end to end in roughly two hours on one RTX 3090, about 3 RMB (~$0.40) of rented GPU time. That shrinks the experiment feedback loop from days to a coffee break.
- Multimodality is stripped to its essential trick: freeze a pretrained vision encoder, learn only the projector and the small LM, and supervise on image-text pairs (ALLaVA-4V). Nothing important is hidden behind a framework abstraction.
- It's the multimodal sibling of the MiniMind LLM project, so the language backbone is itself trained from scratch — you can trace one model from raw text pretraining all the way to answering questions about an image.
Who It's For
Great fit if you want to understand VLM architecture by reading and editing every line, or you teach or study multimodal ML and need a runnable reference that fits a consumer GPU. Look elsewhere if you need state-of-the-art image understanding for a product — at 65M parameters this is a study model, not a deployable one, and its grasp of fine-grained visual detail is bounded by scale and the SFT data.
