Open models, explained as you go.
Four short paths — just enough to make sense of what you’re looking at, then straight to the real data. Every figure below is counted from the catalog, so you can click any of it to see where it came from.
- Start hereOpen models are AI models whose weights you can download, inspect and run yourself.
- Run locallyThree things decide what you can run: memory, quantization and the runtime.
- Understand modelsThe names look chaotic, but they encode a simple structure.
- Build with themMost tools speak the same API, so switching models is usually a configuration change.
Start here
Open models are AI models whose weights you can download, inspect and run yourself.
- 18Permissive licenseFree to use, including commercially
- 5Restricted licenseWeights are downloadable; the license limits some uses
- 2License unknownCheck the source before using it
A model’s “weights” are the billions of numbers learned during training. When a lab publishes them, anyone can run the model on their own hardware — no account, no per-token bill, and your data never leaves your machine.
“Open” covers a range. Some models use permissive licenses like Apache 2.0 or MIT; others allow broad use but restrict certain commercial cases. Mutinai always shows the license so you can tell the difference.
The ecosystem moves quickly: a handful of labs release new families every few months, and a much larger community quantizes, fine-tunes, benchmarks and builds tools around them.
Run locally
Three things decide what you can run: memory, quantization and the runtime.
Llama 3.1 8B, as published
6 precisions on fileMemory is the hard limit. The whole model has to fit in your GPU’s memory (or your Mac’s unified memory) along with a working cache that grows with the length of the conversation. If it doesn’t fit, some runtimes can spill into system RAM — it works, but slowly.
Speed is mostly about how fast memory can be read. That’s why a GPU with fast memory generates text quicker than a CPU with lots of slower RAM — and why mixture-of-experts models, which read only part of their weights per token, feel fast for their size.
Runtimes load the model and serve it: llama.cpp and Ollama run almost anywhere, MLX is built for Apple silicon, and vLLM is designed for serving many users on NVIDIA or AMD GPUs.
Understand models
The names look chaotic, but they encode a simple structure.
- FamilyLlamaA lineage from one developer — Meta.
- ReleaseLlama 3.1A dated generation within it, published 23 Jul 2024.
- ModelLlama 3.1 8BOne trained size: 8.03B parameters.
- VariantLlama 3.1 8B InstructTuned to follow instructions and hold a conversation. The usual choice.
- DownloadQ4_K_MOne file at one precision — 4.6 GiB, published by LM Studio Community.
Read a name from the outside in and it comes apart into five levels. Anyone can add to the lower ones: a quantization, a fine-tune or a distill is often published by someone other than the lab that trained the model, which is exactly what open weights make possible.
Benchmarks give a rough sense of capability, but developers report them with different prompts and settings. Use them to shortlist, then look at community results and reviews from people doing what you want to do.
Mutinai labels every number with where it came from: developer-reported, measured by the community, or estimated.
Build with them
Most tools speak the same API, so switching models is usually a configuration change.
Runtimes like Ollama, llama.cpp’s server and vLLM expose an OpenAI-compatible API. Chat interfaces, coding assistants and agent frameworks can point at them directly — or through a gateway that routes between local and hosted models.
Coding assistants such as Aider and Continue work well with code-specialised models. For agents, look for strong tool-use results and reviews that mention reliability over long tasks.
When a model is close but not quite right, fine-tuning frameworks adapt it to your data on a single GPU. Keep the result’s license — inherited from the base model — in mind.