jingyaogong/minimind-v resource

Health v2 · maintenance only

Adoption not part of the score

8488 stars · 932 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

MiniMind-V is an open-source educational project that trains tiny (26M-200M parameter) vision-language models from scratch in pure PyTorch, with a minimal 65M model trainable in ~2 hours on a single RTX 3090. It includes the full VLM stack—SigLIP2 vision encoder, MLP projector, LLM, plus dataset cleaning, pretraining, and SFT code—serving as both a minimal implementation and a tutorial for understanding multimodal models.

Use cases

• train a vision-language model from scratch on a personal GPU
• learn how VLMs work by reading minimal code
• understand pretraining and SFT for multimodal models
• experiment with tiny multimodal models cheaply
• build an image-understanding chatbot
• study MoE and dense VLM architectures
• fine-tune a small VLM on custom image-text data

When to choose

• you want to learn VLM internals with a transparent, minimal PyTorch implementation
• you have limited compute (single consumer GPU) and want to train a multimodal model end-to-end
• you need a small, OpenAI-API-compatible VLM for experimentation or education

When to avoid

• you need state-of-the-art vision-language performance for production workloads
• you want a battle-tested enterprise multimodal solution rather than an educational codebase
• you need large-scale, high-resolution, or multi-image production features

Facets

learning-resource · maturity active

machine-learning deep-learning llm-training image-processing artificial-intelligence machine-learning computer-vision education python cross-platform vision-language-model vlm multimodal from-scratch siglip pytorch educational small-language-model moe sft pretraining natural-language-processing gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="jingyaogong/minimind-v")