After this explainer you canTell closed, open-weights and open-source models apart, and choose which kind fits a need such as privacy, control or safety.
The short answerSome AI models stay with the company that made them. Others can be downloaded, so anyone can run them and change them.
In simple words
- Closed: you use it through the company’s app or website.
- Open: you can download it and run it yourself.
- Open models help researchers and keep data private.
- But anyone who downloads an open model can remove its safety rules.
The short answerAn open-weights model can be downloaded and run on your own hardware, within its licence. A closed model is used only through apps and APIs that its maker controls or licenses. Truly open-source AI also shares how it was built.
In simple words
- Closed models, such as GPT-6 and Claude, cannot be downloaded; they run only where their maker allows.
- Open-weights models, such as Moonshot’s Kimi K3 and Z.ai’s GLM-5.3, can be downloaded and changed.
- Open weights are not open source, which also shares the code and detailed information about the training data.
- Open models help research and privacy, but their safeguards are easier to remove.
The short answerOpenness is a spectrum: API-only access, open weights under licences of varying freedom, and open-source AI that also releases data information and complete code. Releases cannot be recalled, and EU law treats open-source models differently below the systemic-risk line.
Key points
- Licences range from permissive ones, such as Apache 2.0 or MIT, to custom licences with limits.
- The OSI definition requires data information, complete code and the parameters.
- Cheap fine-tuning, or even direct weight edits, can strip refusals, and a release cannot be recalled.
- EU open-source models skip documentation duties, but not the copyright and summary duties.
| What you can do | ClosedUse it through the maker’s app or API | Open weightsDownload the model itself | Open-source AIModel, code and data details |
|---|---|---|---|
| Use it through an app or API | Yes | Yes | Yes |
| Download and run it yourself | No | Yes | Yes |
| Keep your data on your own computers | Rarely | Yes | Yes |
| Change it on your own computers | No | Yes | Yes |
| Read the complete training code | No | Rarely | Yes |
| Detailed information about the training data | No | Rarely | Yes |
| Maker can fix or switch off every copy | Yes | No | No |
| Examples | Claude Fable 5.2 · Gemini 4 Carbon · GPT-6 Astra | Kimi K3.1 Pro · DeepSeek V4.1 Flash · Inkling 2 Small | Weights, code and data information |
Words to know
- Closed model
- A model that stays with its maker; you use it through an app or API.
- Open weights
- A model you can download, run and change yourself.
- Open-source AI
- Open weights plus the code and details about the training data.
Two ways to share a model
A closed model is like a service you rent: you can use it, but it stays with the company. The company can change it or switch it off.
An open model is like a finished program you can download, copy and change. If your computers are strong enough, you can run it yourself, without sending your data anywhere. Small open models run on a laptop; the biggest need powerful servers.
Many people call these models “open source”. Most are only “open weights”: you get the model itself, but not the code and data details used to build it.
Why it matters
Open models let researchers study AI and let companies keep their data private. But once a model is downloaded, anyone can remove the safety rules built into it.
Many strong open models now come from Chinese companies, and anyone in the world can download them.
Try it yourself
Look at the Examples row in the table above. For each model, decide: could a hospital run it on its own computers?
Check yourself3 quick questions
A hospital wants to use AI on patient notes without sending them to another company. Which kind of model fits best?
A company switches off its closed model. What happens to the people who used it?
Why do some experts worry about strong open models?
Words to know
- Closed model
- A model that stays with its maker; you use it through an app or API.
- Open weights
- A model you can download, run and change yourself.
- Open-source AI
- Open weights plus the code and details about the training data.
Closed: you use it, but you never get it
A closed model cannot be downloaded. You reach it through an app or an API, a connection that lets other software use the model. The API may come from the maker or from a cloud service the maker works with, such as Amazon Bedrock. A few makers also run closed models inside big customers’ own data centres, such as Gemini on Google Distributed Cloud, but nobody can download them.
The maker can change the model, filter it or switch it off. That gives the maker control over misuse. It also keeps the model’s weights secret: the learned numbers inside it, the parameters from explainer 1.
Open weights: download and run
An open-weights model publishes those numbers. Anyone can run it on their own hardware, adapt it or study it, without sending data to the maker.
Many strong open models now come from China, such as Kimi K3, GLM-5.3 and DeepSeek’s models. Alibaba shares some Qwen models too, such as Qwen3.8 Flash Next, but keeps its top model, Qwen3.8 Max, closed.
Open weights are not open source
The Open Source Initiative has a definition of open-source AI. It asks for the weights, the complete code and enough information about the training data to build a similar system. Most models called open share only their weights.
Safety and the law
Once weights are public, safeguards can be removed, so researchers test how hard that is. In the EU, free and open-source models skip some paperwork duties. This does not cover the most powerful models, which the EU says carry “systemic risk” (explainer 5).
From 2027, California bars download sites from knowingly offering models that leave out its required hidden labels (explainer 3).
Try it yourself
Look at the Examples row in the table above. For each model, decide: could a hospital run it on its own computers?
Check yourself3 quick questions
A company calls its model “open source”, but it publishes only the weights. Under the OSI definition, is it open-source AI?
A law firm must not send client files outside its own servers. Which option lets it use a strong model?
Why can’t a lab fix a safety problem in an open-weights model after release, as it could with a closed model?
What exactly is released
An open-weights release publishes the parameter files and usually code to run them. Licences range from permissive ones, such as Apache 2.0 or MIT, to custom licences that limit some uses or users.
Training data and the full training code are rarely released, which limits how far outsiders can reproduce or audit a model.
The Open Source AI Definition
The Open Source Initiative’s definition, version 1.0 from October 2024, asks for the freedoms to use, study, modify and share. It requires the preferred form for making changes: detailed data information, complete code for training and running, and the parameters.
It does not require publishing all training data, but it does require describing the data well enough to build a substantially equivalent system.
Safety research on open weights
Cheap fine-tuning can remove refusal behaviour from open models: one 2023 study undid the safety training of Llama 2-Chat 70B for under $200. Refusals can even be switched off by editing the weights directly, without any training.
Because copies spread beyond the maker’s control, a release cannot be recalled or patched. Supporters answer that open models let outsiders find flaws, run models privately and check makers’ claims. Fully open releases also let others repeat the training.
The legal treatment
Under AI Act Article 53(2), general-purpose models released under a free and open-source licence with public parameters skip the technical documentation duties, unless they have systemic risk. They still need a copyright policy and a public training-content summary.
The licence must let anyone access, use, change and share the model, and the weights, architecture and usage information must be public. Paid or otherwise monetised releases may not qualify.
From January 1, 2027, California’s AB 853 bars generative AI hosting platforms from knowingly offering models that do not add the required provenance data.
Try it yourself
Look at the Examples row in the table above. For each model, decide: could a hospital run it on its own computers?
Check yourself3 quick questions
A lab releases weights under Apache 2.0 with inference code, but no training code and only a one-line data note. Does it meet the Open Source AI Definition 1.0?
An open-source general-purpose model below the systemic-risk line is released in the EU. Which duty still applies?
Why can an open-weights release not be recalled?
Sources
- The Open Source AI Definition (Open Source Initiative)
- AI Act, Article 53 (EU AI Act Service Desk)
- AI Act, full text with recitals (EUR-Lex)
- AB 853 (California Legislature)
- LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B (Lermen and others, 2023, arXiv)
- Refusal in Language Models Is Mediated by a Single Direction (Arditi and others, 2024, arXiv)
- AI model comparison: which models have open weights (Silicon AI News)
- Gemini on Google Distributed Cloud, in customers’ own data centres (Google Cloud, August 2025)
- AI model release tracker: on-device models and cloud availability (Silicon AI News)
This explainer was generated by AI systems under the editorial rules of Silicon AI News and checked against the sources it lists. The live parts come from our checked stories, trackers, model comparison and rules checker.