Open-weights models: AI you can download
An open-weights model is an AI model whose trained parameters, known as weights, are published and can be downloaded. That means you can run it on your own hardware, adapt it or build it into your tools, within the limits of its licence.
Data as of October 22, 2026
Key figures
| Level | Intermediate |
|---|
The definition
At heart, an AI model is a vast collection of numbers: its parameters, or weights, set during training. When those weights stay on the maker’s servers, you can only use the model through its app or API. When they are published, anyone can download them and run the model themselves. That is an open-weights model.
A concrete example
An accounting firm wants to summarise client documents without sending them to an online service. It downloads a mid-sized open-weights model, installs it on an in-house server with inference software, and staff use it like any other assistant. The documents never leave the firm’s network. In exchange, the firm has to manage the hardware, updates and security, and the model it picks will usually trail the best closed models of the moment.
Open weights versus open source
The two terms are often mixed up. In 2024 the Open Source Initiative published an Open Source AI Definition: to earn the label, a system must provide its parameters, but also the full training code and enough information about the data for a skilled person to build a substantially equivalent system. Most so-called open models release their weights without all of that, so “open weights” is the more accurate term for them.
Licences matter too: some allow broad use, including commercial use, while others add conditions. Makers such as Meta, Mistral AI and DeepSeek publish open-weights models, each with its own rules.
Why it matters
Open weights let you keep data in-house, fine-tune a model on your own examples, shrink it through quantization so it fits on modest hardware, and avoid being exposed to a provider’s price or policy changes. The site’s open source models page lists the most downloaded open models, with their licence and size.
Common myths
- “Open means free with no strings attached.” The download is free, but the licence may restrict some uses, and running the model costs money in hardware.
- “Open models are less safe.” Anyone can modify them, which raises genuine questions, but they can also be inspected and audited more freely.
- “They’re for experts only.” Consumer-friendly apps now let you launch a small model on a personal computer in a few clicks.
Related terms
- Model parameters · The internal numbers a model tunes during training; how many it has indicates its size, not directly how good it is.
- Fine-tuning · Continuing the training of an already trained model on a small, targeted set of examples to adapt it to a task, style or output format.
- Quantization · A technique that stores a model's weights with fewer bits (8, 4 or even fewer) so it needs far less memory to run.
- GGUF (model file format) · A single-file format bundling a model's weights, often quantized, with its metadata, so it can run locally with llama.cpp and related tools.
- Small language model (SLM) · A compact language model, light enough to run on a laptop or phone, that performs well on focused tasks rather than everything at once.
- Inference server · Software that loads an AI model into memory and answers requests through an API, while optimizing throughput and GPU memory.
Try it
Full glossary
Frequently asked questions
Is an open-weights model automatically open source?
No. Under the Open Source Initiative's definition, an open source AI system must also provide the training code and detailed information about the data. Many models publish only their weights, sometimes under a restrictive licence.
Do I need a powerful computer to run one?
It depends on the model's size. Small models, especially quantised versions, run on a recent laptop. The largest ones need one or more graphics cards with plenty of video memory, or a rented server.
Does my data stay private with an open-weights model?
If you run it yourself, on your own machine or server, your prompts are not sent to the model's maker. If you use a third-party host, that host's terms apply.
Can I use it commercially?
Often, but not always. Some licences are very permissive, others impose usage conditions or restrictions. Read the licence published with the model before any business use.
Data sources
Data as of October 22, 2026
Report an error
Does a figure, a date or a fact look wrong? Tell us: every report is checked.