Is open-weight the same as open source?

No. Open-weight means the trained parameters are published. The Open Source AI Definition 1.0 also asks for the training code and sufficiently detailed information about the training data, so a release of weights alone does not meet it.

Can I use an open-weight model commercially?

Often yes, but it depends on the licence. Read it for your own case. This is general information, not legal advice.

Do open-weight models send my data back to their maker?

The weights themselves are only numbers in a file and contact no one. The software that runs them might, through update checks, telemetry or web search, so test the whole system with the network cable pulled.

Where can I find open-weight models?

On model hubs such as Hugging Face, and in our own list of open models you can run on your hardware.

ANVE: the private AI office, on the record. See what that means

ANVE

Private AI glossary · updated 2026-10-07

Open-weight model
A model you can download and run in your own building, on its maker’s terms.

Short answer. An open-weight model is an AI model whose trained parameters, the weights, are published so anyone can download them and run the model on their own hardware. Qwen3 and gpt-oss are released under the Apache 2.0 licence, while Llama and Gemma 3 come with their makers’ own licences and use rules. Open-weight is not the same as open source: the Open Source AI Definition 1.0 also requires the training code and detailed information about the training data. For an office, open weights mean the model can run inside the building without calling its maker, but the licence still decides what you may do with it.

What you get and what you do not

What you can do	Hosted model (API only)	Open-weight model	Open source AI (OSI definition)
Download the weights	No	Yes	Yes
Run it on your own hardware	No	Yes	Yes
Training code published	No	Varies, often not	Required
Detailed information about the training data	No	Varies, often not	Required
Rules of use	The provider’s terms of service	Set by the maker, from Apache 2.0 to custom licences	Freedom to use, study, modify and share

Licences in use as of October 2026

The licence is where open-weight models differ most. A permissive licence such as Apache 2.0 sets few conditions. A custom licence can set a size threshold above which you must ask the maker for permission, require a visible notice, or point to a separate list of banned uses. Llama’s, for example, does all three.

Model	Licence	Worth knowing
Qwen3 (for example Qwen3-32B)	Apache 2.0	Listed on the model card
gpt-oss-20b and gpt-oss-120b	Apache 2.0	Listed on the model card
Llama 3.1	Llama 3.1 Community License	A custom commercial licence, read it before you deploy
Gemma 3	Gemma Terms of Use	Refers to a separate Gemma Prohibited Use Policy; Hugging Face asks you to accept the licence before download
Gemma 4	Apache 2.0	Google’s Gemma terms page says Gemma 4 has its own licence, and its licence page shows Apache 2.0

Why it matters for an office

Privacy: the weights are a data file and run offline. Nothing about them needs to contact the maker, so documents can stay inside the building.
Cost: there is no per-token fee. You pay for the hardware, the power and the people who look after it.
Continuity: a model you hold does not vanish when a provider retires it or changes its terms. Keep the exact file and its checksum.
Control: you choose when to update and can test a new version on your own tasks first.
Limits: you cannot inspect the training data, so you cannot know exactly what the model has seen, and the licence may restrict some uses.

What to check before adopting one

Read the licence for your use, including any acceptable-use policy it points to.
Download from the maker’s own organisation page, not a re-upload by an unknown account.
Prefer safetensors or GGUF files, which store only data, unlike pickle-based files that can run code when loaded.
Read the model card for languages, context length, intended use and known limits.
Check that the file fits in memory with room for the context.

Where ANVE stands

ANVE is designed to run open models on a device in the office. In the development build, cloud models are optional, added by an administrator, and their answers are labelled, and every model call is logged with the user, the app, the model and whether it ran locally or in the cloud. On continuity, ANVE’s security page says every order includes a perpetual licence for the software delivered with the unit, that there is no remote kill switch, and that the box keeps working if ANVEAI shuts down. That covers ANVE’s own software; each model keeps the licence its maker gave it.

Questions people ask

Is open-weight the same as open source?: No. Open-weight means the trained parameters are published. The Open Source AI Definition 1.0 also asks for the training code and sufficiently detailed information about the training data, so a release of weights alone does not meet it.
Can I use an open-weight model commercially?: Often yes, but it depends on the licence. Read it for your own case. This is general information, not legal advice.
Do open-weight models send my data back to their maker?: The weights themselves are only numbers in a file and contact no one. The software that runs them might, through update checks, telemetry or web search, so test the whole system with the network cable pulled.
Where can I find open-weight models?: On model hubs such as Hugging Face, and in our own list of open models you can run on your hardware.

What is not real yet

No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.

Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.

Open-weight modelA model you can download and run in your own building, on its maker’s terms.