Short answers to the questions I am asked most about open source AI, grouped by topic, each pointing to the chapter of the Open Source AI Hub that goes deeper. Last reviewed September 2026. Nothing here is legal advice; read the license text for anything that ships.
Definitions
What is open source AI?
Open source AI is an AI system that you can use, study, modify and share for any purpose without asking permission, and whose code, parameters and training-data information are all available under open terms. That is the Open Source Initiative’s Open Source AI Definition, released in October 2024. Most models described as “open” are open-weight rather than open source. Chapter 2 explains the difference.
What is the difference between open weights and open source?
Open weights means the trained parameters are published so you can download and run them; the license may still restrict what you do. Open source means the whole system, parameters, code and data information, is available 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. See chapter 2.
What is the OSI definition of open source AI?
The Open Source AI Definition, version 1.0, published by the Open Source Initiative on 28 October 2024, requires the four freedoms of use, study, modify and share, and names three components that must be available under open terms: data information, code and parameters. It requires detailed information about the training data rather than the data itself, which is its most debated point. Chapter 2 covers the debate.
What is an open-weight model?
An open-weight model is one whose parameters are published for download, so you can run it on your own hardware and fine-tune it. The license can be permissive, like Apache-2.0 on Qwen3, or restricted, like the Llama Community License. Open-weight says nothing about whether the training code or data are available. Chapter 4 lists the families.
What does open data mean for an AI model?
Open data means the training corpus itself is released under an open license, so anyone can inspect what the model learned from. It is rarer than open weights and rarer than open source; Ai2’s OLMo models with the Dolma corpus are the clearest example. A model can be open source under the OSI definition without releasing its data. Chapter 2 places the examples on the spectrum.
What is open washing?
Open washing is describing a restricted or partial release as “open source” to borrow the term’s reputation. The word was popularized in AI by a 2024 ACM FAccT paper and by the Open Source Initiative’s statements that Llama does not meet its definition. The fair test is the license text, not the launch announcement. Chapter 2 keeps it factual.
Licensing
Not legal advice. Read the license file for the exact release you use, and ask counsel for anything that ships to customers.
Is Llama open source?
No, not under the Open Source Definition. Llama models ship under the Llama Community License, which requires a separate license above 700 million monthly active users, incorporates an acceptable use policy, requires “Built with Llama” attribution and a “Llama” name prefix on derivatives, and withholds multimodal rights from EU users. It is a generous proprietary license that permits most commercial use. Chapter 3 reads it clause by clause.
Can I use Llama commercially?
Yes, if your products had fewer than 700 million monthly active users in the month before the model’s release, you follow Meta’s acceptable use policy, and you meet the attribution and naming rules when you redistribute. EU-based companies should check the territory clause on the multimodal models. Details in chapter 3.
Can I use DeepSeek commercially?
Yes for DeepSeek R1 and for the V3 releases from March 2025 onward, which are under the MIT license with attribution as the only obligation. The original December 2024 V3 weights were under a separate DeepSeek model license, so check the revision you downloaded. Chapter 3 has the history.
Is Gemma open source?
No. Gemma ships under Google’s Gemma Terms of Use, which incorporate a prohibited use policy, pass that policy on to anyone you distribute derivatives to, and let Google update the terms. Commercial use is permitted within the policy. Chapter 3 explains the pass-through clause.
What is the difference between MIT and Apache-2.0 for a model?
Both are permissive and allow commercial use, modification and redistribution with attribution. Apache-2.0 additionally grants you an explicit patent license from the contributors and asks you to carry a NOTICE file and mark changed files; MIT asks only that the notice travels with the work. For weights, the patent grant is the practical difference. See chapter 3.
What is a RAIL license?
A Responsible AI License is a permissive grant with a list of prohibited uses attached, which must be passed on to anyone who redistributes the model. BLOOM, StarCoder and the early Stable Diffusion models use variants of it. Because it restricts fields of use it is not open source under the OSI definition, though it allows most commercial use. Chapter 3 covers it.
Are Mistral models open source?
Some are. Mistral’s open line, including Mistral 7B, Mixtral, Mistral Small and Devstral, is Apache-2.0. Other Mistral models are under the Mistral Research License, which is non-commercial, or the Non-Production License, which forbids production use. Check the license on the specific model, never the vendor. Chapter 3 lists the tiers.
Does the EU AI Act exempt open source models?
Partly. Article 53(2) of Regulation 2024/1689 exempts general-purpose models released under a free and open source license from some documentation duties, but not from the copyright policy or the training-content summary, and not at all for models with systemic risk. Whether use-restricted licenses qualify is debated. Chapter 3 has the three-paragraph version.
Does my fine-tuned model inherit the base model’s license?
Yes, and it inherits the terms of every dataset you tuned it on and of any model whose outputs you distilled from. The strictest term anywhere in that chain applies to the result; a permissive base tuned on a non-commercial dataset becomes non-commercial. Keep a provenance file. The checklist in chapter 3 walks through it.
Models
Which open model should I start with?
For a laptop, an 8B-class model from the Qwen3 or Gemma 3 families at 4-bit quantization. For a single 24 GB GPU, Mistral Small or a 27B to 32B model. For a server, Llama 3.3 70B or the large Qwen3 and DeepSeek models. Check the license first, then run AILane to confirm the fit. Chapter 4 is the full map.
What is quantization, and does it hurt quality?
Quantization stores weights in fewer bits, typically 4 or 8 instead of 16, to cut memory and speed up inference. At 4-bit the quality loss is small for chat and larger for exact tasks like code and math. Q4_K_M is the usual balance point and Ollama’s default. Test at the quantization you will deploy. Chapter 4 explains the formats.
What is GGUF?
GGUF is the file format used by llama.cpp, introduced in August 2023, that packages a model’s quantized weights and metadata in one file. The names on GGUF files, such as Q4_K_M or Q8_0, describe the quantization scheme. It is the format Ollama, LM Studio and most local tools use. Chapter 4 decodes the names.
Is gpt-oss open source?
gpt-oss-120b and gpt-oss-20b, released by OpenAI in August 2025, are open-weight models under the Apache-2.0 license, so you may use, modify and redistribute them commercially. OpenAI did not publish training data information, so under the OSI definition they are open-weight with a permissive license rather than open source AI. Chapter 4 places them.
What is a fully open model?
A fully open model publishes its weights, its training code, its training data and usually its training logs, all under open licenses, so that anyone can reproduce and audit it. Ai2’s OLMo and EleutherAI’s Pythia are the reference examples. They are the right choice when you must be able to say exactly what a model learned from. See chapter 2.
Tools
Is Ollama open source?
Yes. The Ollama server and command-line tool are MIT licensed, and the llama.cpp engine underneath is MIT too. Ollama’s newer desktop application is a separate component, so check the repository for the piece you deploy. Chapter 5 lists every runtime’s license.
Is LM Studio open source?
No. LM Studio is a proprietary desktop application that is free to use, including at work. Its command-line tool and SDKs are MIT, but the application itself is not open source. It is a good tool; it just belongs in the proprietary column of a procurement review. Chapter 5 keeps the columns straight.
Is Open WebUI open source?
Not under the OSI definition since 2025. From version 0.6.6 Open WebUI uses its own BSD-based license with a clause requiring its branding to stay visible unless your deployment is small or you have an agreement. That makes it source-available. It works well; read the clause before putting your own logo on it. Chapter 5 has the alternatives.
Is llama.cpp open source?
Yes, llama.cpp is released under the MIT license. It is the C++ inference engine that runs quantized GGUF models on CPUs and GPUs and sits underneath Ollama and several other tools. If you want to understand local inference, it is the place to start. Chapter 5 covers when to use it directly.
Is vLLM open source?
Yes, vLLM is Apache-2.0 licensed and developed as an open project. It is the high-throughput GPU server for serving open models to many users at once, and the usual next step when a single Ollama server stops keeping up with a team. Chapter 5 explains when to switch.
Is LangChain open source?
Yes, the LangChain and LangGraph libraries are MIT licensed. LangSmith, the same company’s observability and evaluation product, is a proprietary hosted service. Most private deployments need less framework than they expect; chapter 5 and my RAG guide say why.
Which open source coding assistants work with local models?
OpenCode, Aider, Continue, Cline, Roo Code, Tabby, OpenHands and Goose all work with local models through Ollama or an OpenAI-compatible endpoint. For agentic work you want a 30B-class code model or larger; for autocomplete, 7B to 14B is fine. Chapter 6 reviews each one.
Is Claude Code open source? Is GitHub Copilot?
Claude Code is proprietary. GitHub Copilot’s chat extension for VS Code became MIT licensed in 2025, but the Copilot service and its models remain proprietary. OpenAI’s Codex CLI is Apache-2.0 and can call open-weight models. An open tool calling a closed model still sends your code to the vendor; chapter 6 untangles the two questions.
Running models locally
Can I run open models on a laptop?
Yes. An 8B model at 4-bit quantization runs in about 5 GB of memory, and a 20B mixture-of-experts like gpt-oss-20b fits a 16 GB machine. Speed depends on your GPU or Apple silicon. Run AILane to see what your machine handles, then follow my Ollama guide.
Do I need a GPU to run open models?
Not for small models: llama.cpp and Ollama run 4B to 8B models on a CPU at usable speed for one person. A GPU or Apple silicon makes everything faster and is needed for larger models and for serving a team. Chapter 4 explains the memory arithmetic; AILane does it for your machine.
How do I run open models inside a company network?
Serve them from a GPU machine inside the network with Ollama or vLLM, put an OpenAI-compatible API in front, add authentication and rate limits at a reverse proxy, and connect editors and chat interfaces to it. I wrote the seven-step version in How to Run LLMs Inside Your Company Network and the Ollama specifics in Ollama for Teams.
Business and legal
Are open models cheaper than APIs?
Sometimes. Open models replace a per-token bill with hardware and operations costs. At low or spiky volume an API is cheaper; at steady high volume owning capacity usually wins, often sooner than expected. The real advantage is choosing the cost shape that fits your usage. Chapter 7 works through it.
Who supports an open model in production?
You do, or a partner you engage. There is no vendor to call. That means someone owns the hardware, the runtime upgrades, the evaluation suite and the pager, and the cost of that is real. It is exactly the work of private AI deployment consulting, and chapter 7 counts it honestly against the benefits.
Is open source AI the same as on-premise AI?
No. Open versus proprietary is about which model you use; on-premise versus cloud is about where it runs. You can run an open model in a cloud and, with the right agreement, a proprietary one on premises. Most private AI deployments combine an open model with on-premise hosting. Chapter 7 covers the model axis and this article the hosting axis.
History and philosophy
What is the difference between free software and open source?
They describe nearly the same licenses for different reasons. Free software, from the GNU project in 1983 and the Free Software Foundation, holds that users are owed four freedoms as a matter of ethics. Open source, adopted as a name in 1998, argues that openness produces better software and sells the method to business. Chapter 1 tells the story.
What does copyleft mean?
Copyleft uses copyright to guarantee freedom: a copyleft license lets you copy, change and redistribute a work on the condition that every redistributed version carries the same freedoms. The GNU General Public License is the best-known example, and Linux is its proof of scale. Permissive licenses like MIT and Apache-2.0 do not have this condition. See chapter 1.
What are the four freedoms?
The four freedoms of the Free Software Definition, numbered 0 to 3: to run the program for any purpose, to study and change it, to redistribute copies, and to distribute modified versions. Freedom 1 requires the source code. The Open Source AI Definition restates all four for AI systems. Chapter 1 explains where they came from.
Who coined the term “open source”?
Christine Peterson of the Foresight Institute proposed “open source” as the name for the movement at a meeting in Palo Alto on 3 February 1998, called after Netscape announced it would release its browser’s source code, and Eric S. Raymond and Bruce Perens founded the Open Source Initiative later that month to steward it. The phrase itself was older: Caldera used “open source code” in a 1996 press release, and there are scattered earlier uses. February 1998 is when it was chosen deliberately as a label. Chapter 1 has the full account.