What Are Open-Weight AI Models?

You might use a chatbot without thinking about who controls the model behind it. Then an update changes how it answers, or you want help with something private you’d rather keep on your computer.

An open-weight model lets you download the pretrained AI and run it yourself. You’ll need software to use it and a computer that can handle it, but you get to keep a working copy. While chatbots get most of the attention, image and video generators can also have open-weight options.

Downloading an AI model doesn’t necessarily give you unrestricted permission or reveal everything that went into making it, though. Sure, you get more autonomy in how you use AI, but the word “open” still leaves a lot unsaid. Here’s what open-weight models are, what you can do with them and what to consider before trying one for yourself.

More from CNET

What are open-weight models and how do they work?

In a chatbot, the AI model is the software that generates the answers. During training, the model learns patterns by adjusting billions of numerical values called neural network weights. When you ask a question, the model uses your prompt and those learned patterns to predict what text should come next, building its answer one piece at a time. 

Developers of closed models, including those behind ChatGPT and Claude, keep their weights private. An open-weight model makes its trained weights publicly available. That means you can download them and run your own copy using compatible software. 

“Whenever you open-weight a model, all the weights are all over the place on the internet, like on Hugging Face and other servers,” Pierre Stock, VP of science at AI company Mistral, tells CNET. “So it’s virtually impossible to cut access to this model.”

With a free model and app, you can chat locally without a subscription or usage fees. More importantly, a fully local setup also keeps your prompts and files off an AI provider’s servers.

Don’t worry if you’re not tech-savvy, because you won’t have to write code or open a terminal to do it. Desktop apps like LM Studio, Ollama and Jan let you download and run open-weight models with a simple install, giving you a ChatGPT-like interface right on your Mac or PC. You still need a computer that can handle your chosen model. Once everything is installed, you can keep chatting offline.

If you know how, you can also adapt the model through further training, where permitted by its license. You don’t have to train an entirely new model from scratch.

Open-weight vs. open-source

People often use these terms interchangeably, but they don’t mean quite the same thing. An open-weight model can also be open-source, but publishing the weights alone doesn’t qualify under the Open Source Initiative’s definition.

Developers must share the code used to train and run it. They must also provide enough information about the training data for a skilled person to build a model that works much the same way. The terms must allow people to use, study, change and share it.

Under the OSI definition, developers don’t have to share the data they used to train a model. The Free Software Foundation says that should be part of the deal, with users free to use and change the data themselves.

Some open-weight licenses limit business use, meaning a model might be free to download but off-limits for an app you want to sell. That kind of restriction would disqualify it under the OSI definition. An open-source license can still require you to keep the same terms when you share a version you’ve changed.

James Martin/CNET

Which models have open weights?

Chinese models currently lead independent AI testing company Artificial Analysis’s intelligence rankings for open-weight models. You may already know about DeepSeek, but there is also Moonshot AI’s Kimi K3 and Alibaba’s Qwen3.8 family, which offer downloadable weights.

Some US alternatives come from the companies behind familiar chatbots. OpenAI offers gpt-oss models separately from ChatGPT, while Google has Gemma models. Meta has Muse Glimmer and added Muse Spark Open alongside its Llama models, and Nvidia has Nemotron. 

You don’t have to stick with the chatbot giants, either. A US-based startup, Reflection, released Beam, whose weights are now available to download. Another US startup, Thinking Machines Lab, offers its model Inkling.

Europe has a contender in France’s Mistral AI. Its downloadable models include Mistral Large 3 and the smaller Ministral 3 family. The newer Mistral Large 4, dubbed Le Chonk, is available now with downloadable weights.

Not every model mentioned here will run on a laptop. Some larger models need server hardware, while smaller versions can run on personal computers or even smaller devices, like phones.

What are the limits?

An open-weight model can still hallucinate when it’s running offline on your laptop. It can also carry biases from its training. Developers may publish weights while keeping training data private, leaving outsiders unable to fully check what went into them.

Performance varies by model and task, so openness alone won’t tell you whether it will beat a closed alternative. 

The freedom to change a model also means people can weaken or remove its safeguards. 

“Guardrails are necessary to open-weighting a model,” Stock says. When I ask how to keep them safe if people can remove those protections after downloading a model, Stock says, “With a lot of work, you can partially overcome the guardrails,” comparing the effort involved to do so to fine-tuning a model yourself.

Once people have downloaded copies, the original developer has limited control over what happens next. A company can cut off someone misusing its online service, and releasing a safer replacement won’t update everyone’s download. 

Yet those same downloads allow independent researchers to examine models more closely and identify faults. Researchers can test changes themselves, rather than depend on the developer to investigate a complaint.

Microsoft and other companies cited that benefit in an open letter defending open weights after reports of potential US restrictions on Chinese models. They argued that keeping these tools available helps security teams defend against attackers using AI.

“When you open-weight a model, you’re essentially enabling the entire ecosystem to progress together,” Stock tells CNET.

Would an open-weight model be useful to you?

Open-weight models can draft emails, summarize documents, analyze local files and help with coding just like cloud models. However, when you run one locally, its speed and ability to handle long documents depend on both the model and your computer’s memory and processing power.

Larger models usually run more smoothly on computers with plenty of memory or a powerful GPU, like a Mac with 16GB or 36GB RAM, though smaller models may work on an 8GB Mac. 

If you already own the necessary hardware, it’s worth considering installing an open-weight model. But if you have to buy a new computer to run them, weigh that cost against a $20-per-month subscription for a cloud model.

You can also use some open-weight models through their developers’ own websites that run them for you, like DeepSeek and Kimi. That’s less work than downloading and running the model yourself, but your prompts go to the provider.

If you want occasional help writing an email and don’t mind using an online service, a hosted chatbot may suit you just fine. ChatGPT, Gemini and Claude all have free plans, so avoiding a subscription doesn’t automatically require running your own AI.

Open weights give you the option to take more control yourself, not to mention keeping sensitive documents on your device. An assistant that keeps working when you don’t have access to the internet sounds great. So does being able to keep a model you like even after its developer retires the online one. 

“You don’t want to live in a world in which someone can cut off access to your model,” Stock says. 

The future also looks promising with open-weight models being installed on smaller devices. Google’s Gemma models include smaller versions for phones and laptops. Mistral has a similar range of models, including specialized ones intended for particular uses and devices.

“We also have models specialized in audio, for instance,” Stock says, “and those can fit literally on the smartwatch.” 

Contributor

Barbara is a tech writer specializing in AI and emerging technologies. With a background as a systems librarian in software development, she brings a unique perspective to her reporting. Having lived in the USA and Ireland, Barbara now resides in Croatia. She covers the latest in artificial intelligence and tech innovations. Her work draws on years of experience in tech and other fields, blending technical know-how with a passion for how technology shapes our world. See full bio