Every time I show someone a laptop answering prompts with the network adapter disabled, the second question is the same: "wait, are you allowed to do that?" It's a fair question and the answer has more structure than a yes, so here it is in the form I'd have wanted before I started.
Open weights, in one paragraph. Using a hosted model is ordering a meal — you get the output, the recipe stays in the kitchen. An open-weight release publishes the trained parameters, a large file of numbers, and you run it yourself. These are not leaks. The largest labs in the field publish them deliberately.
What the licences actually say.
Google's Gemma, Alibaba's Qwen and Mistral's models ship under Apache 2.0 — use, modify, redistribute, commercialise. OpenAI's open model, gpt-oss, is also Apache 2.0 with a usage policy layered on top. Meta's Llama family uses its own community licence, which is permissive for most purposes but carries extra conditions worth reading if you're at scale.
The practical rule: **read the licence on the model page you download from. **Not a summary, not a blog post, not the licence of the base model the thing was fine-tuned from. Variants and community re-uploads do not reliably inherit the original terms, and quantised conversions are re-uploads.
Fine-tunes and abliteration. Two techniques produce models that stop refusing. A fine-tune retrains on examples that don't refuse — Dolphin and Hermes are the well-known ones. Abliteration is different: researchers found the refusal behaviour occupies an identifiable direction in the model's activation space and can be removed surgically. Neither technique is illegal in the US, UK, EU or Australia. Both are, as far as I can tell, squarely within the base licences.
There is a real engineering cost to abliteration that gets undersold: it is surgery, and surgery has side effects. Abliterated models typically lose measurable ground on reasoning benchmarks and can behave oddly on unusual prompts. Uncensored is not a synonym for better. If a permissive mainstream model with your own system prompt does the job, that's the higher-quality path.
The EU AI Act, since it comes up every time: it places obligations on providers putting AI models and systems on the market, and it treats open-source releases more lightly than proprietary ones. It is not aimed at an individual running a model locally. In the US there's still no comprehensive federal statute and the live arguments are about regulating large providers and about state-versus-federal authority — not about what's on your SSD.
The part that actually matters, and the part most write-ups skip.
The model being legal does not make the output legal. Defamation is defamation. Harassment is harassment. Non-consensual imagery of real people, anything involving minors, material assistance with a crime — all of that is exactly as illegal as it would be if you'd typed it yourself. Running locally changes who can observe you. It changes nothing about what the law permits.
That distinction is the whole thing. "Uncensored" means the tool stops treating you as a suspect by default. It does not mean the rules stopped existing, and anyone selling it to you as though it does is selling you a problem.
A short checklist if you're setting this up:
- Read the LICENSE in the exact repo you pull from.
- Check who uploaded it and how many downloads it has.
- File format should be GGUF or safetensors. An .exe from a random site calling itself an "uncensored AI installer" is malware wearing a costume.
- Prefer Apache 2.0 or MIT when quality is close — in this category it increasingly is.
- If you're outside the US, UK or EU, spend two minutes on your local rules. Some jurisdictions are stricter.
I'm not a lawyer and this isn't legal advice. It's the read I did before shipping a product in this space, written down so the next person doesn't have to repeat it.
https://getpocketcore.com/?utm_source=devto&utm_medium=article&utm_campaign=nov5_article11
Top comments (0)