Model training tools and resources

A complete book on building AI agents, published free and in the open by Bojie Li, organised around one formula: an agent is a language model plus context plus tools.

Find college-level computer science courses with video lectures, sorted by topic.

This is a hands-on tutorial that walks through building a language model entirely from scratch in PyTorch, with no shortcuts through libraries like transformers, trl or peft that normally do the heavy lifting.

Higgsfield is an open-source framework for training very large neural networks — the kind with billions to trillions of parameters, such as large language models — across many computers at once, a process the project's own documentation describes as normally involving a lot of manual, error-prone setup ("multi node training without crying").

The Python library that defines how open models are written down, covering text, images, audio, video and combinations of them, for both running a model and training one.

Train your own small chatbot from scratch on a GPU server, then chat with it in the terminal.

Load a big open language model from Hugging Face and generate text on a GPU with only a few GB of memory.

Learn to build AI systems yourself, from the math behind models to retrieval pipelines and agents, by writing and running real code.

A small model built to make decisions rather than write text.

MiMo-V2.6-Distill-Qwen-9B is a 9-billion-parameter model made by fine-tuning Qwen3.5-9B on data generated by Xiaomi's larger MiMo-V2.6 models, a technique called distillation where a smaller model is trained to imitate a bigger one's outputs rather than being trained on raw internet text from scratch.

Give one model text, images, video or audio, with up to a million tokens of context such as a whole code repository or a long tool log, and get text back for coding and agent work.

MiMo-V2.6-Pro-RL is Xiaomi's flagship model in the MiMo-V2.6 family, reading and generating text, images, video and audio in one model with a 1-million-token context for handling large codebases or long multi-session agent work.

Control already-trained reinforcement-learning policies for Pollen Robotics' Microduck robot entirely client-side: physics run through MuJoCo compiled to WebAssembly and the policy network runs via onnxruntime-web, with no server involved.