What is an open-source AI model? A plain-English guide
An open AI model is one you can download and run yourself. Learn what open weights and open source mean, what you can do with these models and how to start safely.
Short answer. An open AI model is one whose trained “weights” are published, so you can download it and run it on your own computer or server instead of calling a company’s API. Most are open weights rather than strictly open source, and the license decides what you may do with them, including whether you can use them in a business.
What an AI model actually is
A language, image or speech model is, physically, a very large file of numbers plus a small amount of code that knows how to use them. The numbers are called weights, or parameters. They were adjusted during training, when the model processed enormous amounts of text, images or audio until it became good at predicting what comes next. After training the weights are fixed, and using the model to get answers is called inference.
Model sizes are quoted in parameters. A “7B” model has 7 billion of them, and the largest open models in our catalog have more than a trillion. Size matters because the weights must fit in memory while the model runs, which is the subject of how much memory you need.
What “open” means
With a closed model, such as the ones behind the big chatbot apps, you never get the weights. You send your text to the company’s servers and get an answer back, under the company’s terms and price list.
With an open model, the developer publishes the weights, usually on Hugging Face, together with a license. You can:
- run the model on your own hardware, even offline;
- keep your data on your own machine;
- adapt the model to your needs through fine-tuning;
- build it into a product, if the license allows it;
- study how it behaves.
“Open” is not one standard, though. Most releases are open weights: the numbers are public, but the training data and the training code usually are not. Some licenses are fully permissive, and others add conditions such as revenue limits or credit lines. Open weights vs open source explains the difference, and AI licenses explained shows how to read the terms.
Open models vs closed models behind an API
| Open-weights model | Closed model behind an API | |
|---|---|---|
| Where it runs | Your hardware, or a host you choose | The vendor’s servers |
| Your data | Can stay on your machine | Is sent to the vendor |
| Cost | Hardware and electricity, or hosting fees | Pay per use or by subscription |
| Customizing | Fine-tuning, if the license allows it | Only what the vendor offers |
| Upkeep | You handle setup and updates | The vendor handles it |
| Rules | Set by the license you downloaded | Set by terms the vendor can change |
Neither is better in every case. Open models are attractive when privacy, cost at scale, offline use or control matter. Closed models are attractive when you want the least setup. Many teams use both.
The kinds of open models
Open models are not only chatbots. The catalog on this site covers five kinds:
- Language models answer questions, write and edit text, reason and code, for example Gemma 4, gpt-oss and Qwen3.8-27B. Some do one narrow job, such as reading scanned pages (OCR) with DeepSeek-OCR or translating with Hy-MT2.
- Image models create and edit pictures, for example FLUX 3 Image Klein and Z-Image.
- Video models create short clips, for example Wan 2.2 and LTX-2.5.
- Speech models turn speech into text or text into speech, for example Whisper large-v3-turbo, Qwen3-TTS and Kokoro.
- Embedding models turn text, images or audio into numbers so that software can search by meaning, for example Qwen3-Embedding.
Each kind has its own memory needs, and how to choose an open AI model shows where to start with each. Browse them all in the model catalog.
What open models do not guarantee
- Not always free for business. The download costs nothing, but some licenses limit commercial use, large companies or certain applications.
- Not always small. The biggest open models need server racks. The guide on memory shows how to tell what fits on your machine.
- Not automatically accurate or safe. Every model can make mistakes, including hallucinations. Treat outputs as drafts to check.
- Not always transparent. Training data is usually not published, and the benchmark scores in announcements come from the developers themselves.
How to get started
- Check your hardware. The memory of your graphics card, or of your Mac, decides what you can run. The VRAM calculator and the models by VRAM page help.
- Pick a model. Start from a shortlist such as the best open-source LLMs or the small models for laptops, or from the matching list for speech, documents or search.
- Run it. Ollama and LM Studio are the easiest first tools, and how to run AI models locally walks you through them.
- Read the license before you use the model for work. The page on models you can use commercially groups them by license type.
Frequently asked questions
- Is an open-source AI model free?
- The download is usually free, but running a model costs hardware, electricity or hosting fees, and the license may limit commercial use. Check the license before you build a product on it.
- Can I run an open AI model offline?
- Yes. Once the weights are downloaded, tools such as Ollama, LM Studio and llama.cpp run models on your computer without an internet connection.
- Are open models as good as the closed models behind popular chatbots?
- Some of the largest open models are competitive with leading closed models on benchmarks, but those scores are mostly published by the developers themselves, and models small enough for a laptop are less capable than the largest ones. Test a shortlist on your own tasks.
- Is it safe to download open models?
- Download from the developer's official page and prefer file formats such as safetensors and GGUF, which are designed to hold only data. Be careful with any model that asks you to run extra code.
- Do I need a powerful computer to use an open model?
- Not always. Small models run on an ordinary laptop with 16 GB of memory, while the biggest open models need servers with many GPUs. The memory a model needs is listed on every model page.
Key terms in this guide
- Fine-tuning Fine-tuning continues a pretrained model's training on your own examples so that it behaves better for a specific task, style or domain.
- Inference Inference is the act of running a trained model to get an output, such as answering a prompt, as opposed to training it.
- Open-source AI Open-source AI, as defined by the Open Source Initiative, is an AI system you may use, study, modify and share freely, with the weights, the code and detailed information about the training data all available.
- Open weights An open-weights model is an AI model whose trained parameters, called weights, are published so that anyone can download and run it, usually under a license that sets the terms of use.
- Parameters Parameters are the numbers inside a model that are adjusted during training; their count, such as 7B or 70B, is the usual measure of a model's size.
- Permissive license A permissive license, such as Apache 2.0 or MIT, lets you use, modify and sell a model with very few conditions, usually just keeping a notice.
Models mentioned in this guide
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| gpt-oss OpenAI | 21B – 117B 5.1B active | 128K | Apache 2.0 (Permissive) | 15 GB Laptop or phone | |
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Olmo 3.1 Allen Institute for AI (Ai2) | 7.3B – 32.2B | 64K | Apache 2.0 (Permissive) | 5.04 GB Laptop or phone | |
| FLUX 3 Image Klein Black Forest Labs | 5.6B | – | FLUX Kommunity (Conditional) | 8 GB Laptop or phone | |
| Whisper large-v3-turbo OpenAI | 810M – 1.54B | – | MIT (Permissive) | 6 GB Laptop or phone |