I installed Red Hat Linux in June 1998, four months after a small group in Palo Alto adopted “open source” as the name for what was already happening. The two words had been used before; February 1998 is when they became a movement. What drew me in was not the software but the philosophy behind it, the idea that people should be free to study, change and share the tools they depend on. I became an open source lover first and a programmer later: it stayed a hobby until my professional career began in 2000, and it has been part of my production work ever since. In April 2006 I wrote on this blog that open source was a hit in India. Today most of my private AI work runs on open-weight models served with Ollama or vLLM, and on open tools like OpenCode, inside networks that never touch a public API. Clients on OpenAI, Anthropic, AWS, Google and Azure are a separate part of my practice, and that is where my open source AI consulting stays honest about when open is not the answer. This hub collects what I have learned about the open side of AI: what the words actually mean, how the licenses work, which models and tools are worth your time, and when choosing open is the right call.
It also corrects a habit I had myself. For two years this site used “open source” and “open-weight” as if they were the same thing. They are not, and the difference has practical consequences for what you can ship. From here on the hub uses the terms precisely: an open-weight model is one whose parameters you can download; an open source AI system is one that meets the Open Source Initiative’s definition, which asks for more than weights.
What this hub covers
Three layers, in the order I think about them. Theory first: where open source came from, what it demands, and how the definition extends to AI. Then the landscape: the model families, the tooling stack, and the coding assistants that work with local models, each with its license stated. Then the decision: why open models win for some workloads and lose for others. Practice is the product of theory here. You cannot read a model license well if you do not know why the Open Source Definition forbids discrimination against fields of endeavor, and you cannot pick a runtime well if you do not know what a quantized weight is.
Start here
- Engineering leaders and decision-makers: read what open source AI means, then why choose it, and when not to, then the license comparison, then the model landscape.
- Engineers who will run this: read the definitions, then the models, the toolbox and the coding assistants, then the operational guides listed below.
- If you want the story: start with free software vs open source. It explains why every argument about “open” AI is an argument we already had in 1998.
Chapters
- Free Software vs Open Source: The History and Philosophy Behind the Licenses: GNU, the four freedoms, copyleft, the 1998 birth of “open source”, and why the ethics-versus-pragmatism split matters for AI.
- What Is Open Source AI? Open Weights vs Open Source vs Open Data: the OSI’s Open Source AI Definition, the openness spectrum from OLMo to Llama to closed, and open washing.
- Open Model Licenses Compared: Apache-2.0, MIT, the Llama Community License, Gemma Terms, RAIL, Mistral’s tiers, the EU AI Act, and a checklist to run before you ship.
- The Open Model Landscape: the families by maker and by task, sizes and quantization explained, and how to choose one for your hardware.
- The Open Source AI Toolbox: runtimes, chat UIs, RAG frameworks, vector stores, fine-tuning, evaluation and observability, with every tool’s license.
- Open Source AI Coding Assistants: OpenCode, Aider, Cline, Continue, Roo Code, Tabby, OpenHands and Goose, and how they compare with Claude Code, Copilot and Codex.
- Why Choose Open Source AI, and When Not To: the honest case for open models against the real trade-offs.
- Open Source AI FAQ: short answers to the questions people actually ask, from “Is Llama open source?” to “Is LM Studio open source?”
Run it yourself
This hub is the what and the why. The how is already written up elsewhere on this site, from the same client work:
- How to Run AI Models Locally and in the Cloud with Ollama: the first model on your own machine.
- AILane: my free tool that tells you which open-weight models your hardware can actually run.
- How to Run LLMs Inside Your Company Network and Ollama for Teams: serving models to a whole team.
- How to Build a Private RAG System: open models over your own documents.
- Private AI Architecture Checklist: what to audit before it goes live.
- Building an Education Marketplace with Private AI Code Generation: a real project delivered with OpenCode and Ollama.
Why it matters
Open source once made IT a ladder that anyone with a second-hand computer could climb. AI has raised that floor, and I argued in Who Cannot Use AI, and Why? that open-weight models, efficient small models and on-premise deployment are what keep the door open. There is a harder-nosed reason too: open models are what make private AI possible at all. If the weights are not yours to run, the question of on-premise versus cloud never even arises.
Working with me
Everything in this hub is free to read and use. The same knowledge is what I bring to open source AI consulting: choosing a model and a license you can ship, sizing the hardware, and running it inside your own network. If you would rather have it designed, deployed and supported than read about it, see private AI deployment consulting, my consulting practice, or tell me about your situation.
Frequently asked questions
What is open source AI?
Open source AI is an AI system you can use, study, modify and share for any purpose without asking permission, with the code, the model parameters and sufficiently detailed information about the training data all available under open terms. That is the Open Source Initiative’s definition. Most models marketed as “open” are open-weight rather than open source: you get the parameters, but not the rest, and often not the freedom to use them for anything.
Is Llama open source?
No, not under the Open Source Definition. Llama ships under the Llama Community License, which caps use at 700 million monthly active users without a separate agreement, incorporates an acceptable use policy, and imposes naming and attribution rules. It is a generous proprietary license that permits most commercial use. Details in the license comparison.
What is the difference between open weights and open source?
Open weights means the trained parameters are published for download. Open source means the whole system meets the OSI definition: parameters, code and data information, all under a license that does not restrict who may use it or for what. Every open source model is open-weight; very few open-weight models are open source. The definitions chapter walks the spectrum.
Can I use open models commercially?
Usually, but it depends on the license of the exact model you downloaded. Apache-2.0 and MIT models such as Qwen3, DeepSeek R1 and gpt-oss allow commercial use with attribution. Llama and Gemma allow it within their use policies and thresholds. Research-only licenses, such as some Mistral releases, do not. Read the license file, not the model card summary.
Is Ollama open source?
Yes. The Ollama server and command-line tool are released under the MIT license, and it runs on top of llama.cpp, which is also MIT. The newer desktop application is a separate question, so check the repository for the component you are shipping. The toolbox chapter lists the license of every tool I recommend.
Where do I start if I want to run an open model today?
Install Ollama, run AILane to see what your machine can handle, and pull a model that fits. My Ollama guide takes you from install to a working local API in an afternoon. For a team, read how to run LLMs inside your company network.