Open-weight AI models: what you can run locally, and what ‘open’ does not promise

Open-weight models let you download a model and run it on your own computer, potentially offline and without sending prompts to the provider. But downloadable weights do not automatically mean open-source software, unrestricted use or a model that will run happily on any laptop.

Watch Desk analysis

What happened

CNET’s guide to open-weight AI models explains that a model’s weights are the learned numerical values it uses to generate answers. If those weights are publicly available, users can download a copy and run it with compatible software. Desktop apps such as LM Studio and Ollama can provide a chat-style interface without requiring users to write code or open a terminal.

A local setup can keep prompts and files on your device, and can continue working offline once installed. CNET also notes that open-weight options exist for image and video generation, not only chatbots. Its examples include models from DeepSeek, Alibaba, OpenAI, Google, Meta, NVIDIA, Mistral and others; not all will run on a personal computer.

Why it matters

The useful choice is not simply “open” versus “closed”. Running a model locally can offer more control over sensitive material and access to a copy even if the provider retires an online service. The trade-off is that your computer needs enough memory and processing power, and performance depends on both the model and the task. If the hardware is already on your desk, local use may be worth exploring; buying a new machine just to avoid a cloud subscription is a different calculation.

There is a further distinction worth keeping straight: making weights available does not by itself meet the Open Source Initiative’s definition of open source. CNET says that definition also calls for training and running code, sufficient information about training data, and terms allowing people to use, study, change and share the software. Licences can also restrict commercial use, so “free to download” is not the same as “free to build a business on”.

Our read

This is a useful map of the options, not a promise that local AI is automatically private, safe or better. Models can still hallucinate and inherit bias; developers may not disclose training data, and a downloaded copy can be modified after it leaves the developer’s control. Read the licence, check your machine’s capacity, and be clear about whether you need offline access or simply want a free chatbot. “Open” is a starting point for questions, not the answer to all of them.

What to watch

• Whether the model’s licence permits your intended personal or commercial use.
• Whether your computer can run it at a usable speed and handle the work you want to do.
• What data stays local, and what is sent to a provider if you choose a hosted version.

Discussion spark: For everyday AI use, would you choose a local model for greater control over your data, or a hosted service for convenience and less hardware hassle?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.