Open-weight model
Also known as: open-weights model, open weight model, open-weight LLM
An open-weight model is an AI model whose trained parameters, or weights, are released publicly so anyone can download, run, and often fine-tune it on their own hardware.
That differs from closed models available only through a company's API. While open-weight models are often called open source, many don't share their training data or code, and their licences may restrict commercial use, so 'open' covers a wide range.
Open-weight model, explained at three levels
At home
Think of a closed AI model as a restaurant meal: you can only have it at the restaurant, on their terms. An open-weight model is like the restaurant handing out the finished dish to take home, where you can reheat it, change the seasoning, and serve it however you want. You download the model's files and run it on your own computer, without sending your conversations anywhere.
At work
An open-weight model is an AI model whose trained weights are published for download, so organizations can run it on their own servers or a chosen cloud instead of using a vendor's API. That gives more control over data location, cost, and customization. The training data and code usually stay private, and the licence decides which commercial uses are allowed.
At home
A model's 'weights' are the billions of numbers it learned during training, stored in files that can be several to hundreds of gigabytes. Apps such as Ollama and LM Studio download those files and run them with an engine like llama.cpp. Publishing weights is like sharing a finished cake: you can eat it, slice it, or add icing, but you don't get the recipe or shopping list, meaning the training code and data.
At work
Licences vary widely. Some models use permissive open-source licences, such as Apache 2.0 for Mistral 7B and OpenAI's gpt-oss, and MIT for DeepSeek-R1. Others, such as Meta's Llama licence, allow most commercial use but add conditions like user-count thresholds and acceptable-use rules. Teams self-host with serving engines such as vLLM or llama.cpp, or use cloud platforms that host open models, and often fine-tune them with LoRA for their own tasks.
At home
Safety training lives in the same weights as everything else, so once those weights are public, anyone can fine-tune the refusals away for a modest amount of computing. That's a real risk, and a release can't be recalled. While closed models can be patched centrally, open weights spread research, let small companies compete, and let you inspect what you run. In my opinion, that benefit outweighs the risk, because it stops a few giants from controlling AI.
At work
Open weights give architectural freedom: air-gapped deployments, data residency in Canadian data centres, fixed versions that never change under you, and freedom to fine-tune. In exchange, you own the operations: GPUs, scaling, patching, and monitoring. Evaluation should cover licence terms, provenance and country of origin, built-in censorship or bias, and file security, since some model formats can carry malicious code, so prefer safetensors files. At low volume, self-hosting often costs more than API fees.
Real-world examples
At home
Thomas downloaded an open-weight model and ran it on his gaming PC. He could chat with it offline and knew none of his conversations were sent to a company's server.
At work
Lena is an IT director for a Canadian municipality. She chose an open-weight model hosted in a Canadian data centre so residents' data stayed in the country and under the city's control.
How it compares
| feature | Open-weight model | Closed model |
|---|---|---|
| Weights available | Yes, downloadable | No, API only |
| Training data shared | Usually not | No |
| Self-hosting | Possible | Not possible |
| Top capability | Close behind leaders | Often the leaders |
Open-weight models give control. Closed models often lead on raw capability.
For analysts
Before adopting an open-weight model, I'd have someone read the licence carefully. 'Open' can still mean usage limits that matter for a commercial product.
History & origin
Meta released its LLaMA model to researchers in February 2023, and the weights leaked publicly within about a week. Meta then released Llama 2 in July 2023 under a licence allowing most commercial use but restricting very large companies. France's Mistral AI released Mistral 7B under the permissive Apache 2.0 licence in September 2023. Toronto-based Cohere published weights for its Command R models for non-commercial use in 2024. The Open Source Initiative published its Open Source AI Definition in October 2024, setting a higher bar that most open-weight models didn't meet. DeepSeek's R1, released in January 2025 under the MIT licence, and Alibaba's Qwen models intensified competition, and OpenAI released its first open-weight models since GPT-2, gpt-oss, in August 2025. In my view, open-weight models are the most important check on a handful of companies controlling AI.