Open-Weight AI Models Explained: What You Can Download, Run and Trust

An open-weight AI model lets anyone download a working copy of the trained neural network and run it on their own machine, with no subscription and no prompts sent to a provider's servers. That is the core promise behind a growing slice of the AI market, and it's reshaping who actually controls the technology that answers your questions.

The distinction matters right now. Chatbots like ChatGPT and Claude keep their underlying weights private, so the developer controls every update and every log of your conversation. Open-weight models flip that arrangement. You keep the model. It keeps working even if the developer pulls the plug online.

"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 French AI company Mistral, tells CNET. "So it's virtually impossible to cut off access to this model."

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

In a chatbot, the model is the software that generates answers. During training, it adjusts billions of numerical values called neural network weights until it learns patterns that let it predict what text should come next. Closed-model developers, including those behind ChatGPT and Claude, keep those weights private. An open-weight model publishes them, so anyone can download the weights and run a personal copy with compatible software.

The practical payoff is twofold. A free model plus a free app means local chatting with no subscription or usage fees. It also means your prompts and files stay on your device rather than landing on an AI provider's servers.

And you don't need to be a programmer. Desktop apps like LM Studio, Ollama and Jan let users download and run open-weight models with a simple install, delivering a ChatGPT-like interface on a Mac or PC. The catch is hardware: you still need a computer that can handle your chosen model. Once everything is installed, the chat works offline.

Users with the skills can go further and adapt a model through additional training, where the license permits it — no need to train a new model from scratch.

Open-weight vs. open-source: what's the difference?

People often use the terms interchangeably, but they don't mean the same thing, and the gap has real consequences for what you're allowed to do.

Under the Open Source Initiative's definition, publishing weights alone doesn't qualify as open-source. Developers must also share the code used to train and run the model, and provide enough information about the training data for a skilled person to build a comparable model. The terms must allow people to use, study, change and share it.

The OSI definition doesn't require sharing the actual training data. The Free Software Foundation disagrees, arguing the data should be part of the deal, with users free to use and change it themselves.

Licensing adds another layer. Some open-weight licenses restrict business use, meaning a model can 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. Open-source licenses can still require you to keep the same terms when you share a modified version.

Which models have open weights?

Chinese models currently lead independent AI testing company Artificial Analysis's intelligence rankings for open-weight models. DeepSeek is the best known, but Moonshot AI's Kimi K3 and Alibaba's Qwen3.8 family also offer downloadable weights.

US alternatives include:

• OpenAI — gpt-oss models, offered separately from ChatGPT
• Google — Gemma models, including smaller versions built for phones and laptops
• Meta — Muse Glimmer, plus Muse Spark Open alongside its Llama models
• Nvidia — Nemotron
• Reflection — a US startup whose model Beam now has downloadable weights
• Thinking Machines Lab — another US startup, offering its model Inkling

Europe has a contender in 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.

Hardware requirements vary widely. Larger models need server hardware. Smaller versions run on personal computers, and some are compact enough for phones.

What are the limits?

Openness doesn't fix the familiar problems. An open-weight model can still hallucinate while running offline on a laptop, and it can carry biases from its training data. Developers may publish weights while keeping that data private, leaving outsiders unable to fully verify what went into the model.

Performance also varies by model and task. Openness alone won't tell you whether a given model beats a closed alternative.

Then there's the safeguard problem. The freedom to change a model means people can weaken or remove its guardrails.

"Guardrails are necessary to open-weighting a model," Stock says. Asked how those protections hold up once people can strip them after downloading, he answers: "With a lot of work, you can partially overcome the guardrails," comparing the effort involved to fine-tuning a model yourself.

Once copies are downloaded, the developer's control largely ends. A company can cut off someone misusing its online service, but releasing a safer replacement won't update everyone's download.

That same irreversibility cuts both ways. Downloads let independent researchers examine models closely and identify faults, testing changes themselves rather than depending 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. But local performance depends on both the model and your machine's memory and processing power.

Larger models run more smoothly on computers with plenty of memory or a powerful GPU — think a Mac with 16GB or 36GB of RAM — though smaller models may work on an 8GB Mac. If you already own the hardware, installing an open-weight model is worth considering. If you'd have to buy a new computer, weigh that cost against a $20-per-month subscription for a cloud model.

There's a middle path, too. Some open-weight models are available through their developers' own hosted websites, like DeepSeek and Kimi. That's less work than running the model yourself, but your prompts go to the provider.

And avoiding a subscription doesn't require running your own AI at all. ChatGPT, Gemini and Claude all have free plans, so occasional email help through an online service costs nothing.

Why the control question is the whole story

The real argument for open weights is autonomy. Sensitive documents stay on your device. The assistant keeps working without internet access. And a model you like keeps running even after its developer retires the online version.

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

The trajectory points smaller. Google's Gemma line includes versions built for phones and laptops, and Mistral has a similar range of specialized models aimed at particular uses and devices.

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