Some AI makers let anyone download their models. What's in the download is the weights, the numbers set during training, usually without the data or code behind them and always under a licence. Who releases them and why, what open weights is and isn't, why it matters to you, and the risk on the other side.
Meta spent more than 39 million hours of chip time training its Llama 3.1 models, then let anyone download the result for free. What you get is the weights, billions of numbers set during training, drawn here as tiny dials. Whoever has the weights has the model.
The video sets open weights against the closed models behind ChatGPT, Claude and Gemini, shows what a download actually contains and what it leaves out, and explains why open weights isn’t the same as open source. It covers who releases open weights, the licences that come with them, Meta’s own reasons for giving its models away, what it means for your data and your costs, and the other side: once weights are out they can’t be called back, and their safety training can be stripped away.
Every claim was checked against the Llama 3.1, DeepSeek R1 and gpt-oss model cards, the Open Source Initiative’s definition, Meta’s and OpenAI’s announcements, the Llama licence, the US government’s 2024 report on open model weights and research on removing safety training, in October 2026. This is an independent explainer, not affiliated with any AI company.
Transcript
Meta spent more than thirty nine million hours of computer chip time training its Llama three point one models. Then it let anyone download the result for free.
The download is the model's weights, billions of numbers set during training. You can picture each one as a tiny dial. They're what training leaves behind, so whoever has the weights has the model.
Llama's largest version arrives as a hundred and ninety one files of numbers, about eight hundred gigabytes in all. Beside them sit a small settings file and the licence.
ChatGPT, Claude and Gemini work differently. Their makers keep the weights on their own servers, and only your words and the answers travel between you.
Open weights means the numbers themselves are handed over. The model runs on hardware you choose, and nothing you type has to leave it.
Meta isn't alone, and DeepSeek released its R1 model in twenty twenty five. OpenAI released two models it calls G P T O S S, its first open weights models since GPT two. Google's Gemma four followed in April twenty twenty six.
What you usually don't get is the training data or the code that set those numbers. You can run the model and change it, but you can't rebuild it or check what it learnt from.
That's why open weights isn't the same as open source. The Open Source Initiative's definition also asks for the training code and a detailed account of the data. The Allen Institute's model, OLMo, meets it, and publishes its training data as well.
Every download comes with a licence, and they differ. DeepSeek's R1 uses the MIT licence, which allows almost anything. Meta's Llama licence comes with rules, and a company with more than seven hundred million monthly users must ask Meta first.
In twenty twenty four, Mark Zuckerberg wrote that selling access to AI models isn't Meta's business, so giving them away doesn't cost it revenue. He added that companies want to keep their data private, and want models no supplier can change or withdraw.
That choice can change with each model. Meta's newest, Muse Spark, launched in April twenty twenty six for chosen partners only, and Meta says it hopes to open future versions.
For you, open weights matter in four ways. Your data can stay on your own machines, and you can train the model further on your own work. Nobody can change or withdraw your copy, and the cost is your own hardware rather than a bill for each request.
The same freedom goes to everyone. Once weights are out they can't be called back, and researchers stripped the safety training from Llama models for less than two hundred dollars. A US government report in twenty twenty four still advised watching the risks rather than restricting releases.
Open weights means you get the numbers that make the model, without the data or the code that made them, and under a licence worth reading.
Found this useful?
Subscribe for the next one, or tell me what you want explained. I take requests.
