Open-Weight AI Models Explained: Open Weights vs Open Source

In short: an open-weight AI model is one whose trained weights - the billions of numbers it learned in training - are published so anyone can download and run it on their own hardware. That is not the same as open source: the training code and data are often not released, and the licence can restrict how the model is used. Licences range from permissive ones such as MIT and Apache 2.0 to custom licences written for one model. Many of the models ranked on this site publish their weights, from vendors including DeepSeek, Moonshot AI, Z.ai, Xiaomi, Meta and NVIDIA.

What does "open weight" mean?

"Open weight" means the model file itself is public. An AI model is mostly a very large set of numbers, called weights or parameters, adjusted during training until the model gives good answers. When a vendor publishes those numbers, you can download the model, run it on your own computer or server, and change it by further training (fine-tuning).

When a model is not open weight - often called a closed or API-only model - you can only use it through the vendor's app or API, and the weights never leave the vendor's servers.

For what weights and parameters are, see how LLMs work.

Open weight vs open source: what is the difference?

Open weight means the weights are published; open source, as the Open Source Initiative defines it for AI, means much more is published too. The Open Source Initiative's Open Source AI Definition says that the preferred form for modifying an AI system must include three elements:

Element	What the OSI definition requires	Usually published by an open-weight model?
Parameters	"The model parameters, such as weights or other configuration settings"	Yes - this is what makes it open weight
Code	"The complete source code used to train and run the system"	Often only the code to run it
Data information	"Sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system"	Rarely in full

So most open-weight models let you use and adapt the model, but not rebuild it from scratch. That is why the two terms are kept apart.

What licences do open-weight models use?

There are two broad kinds: standard open-source licences, and custom licences written for a model.

• Standard licences. MIT and Apache 2.0 are short, widely used software licences that allow commercial use, change and redistribution, with conditions such as keeping the copyright notice. Several models in the table below use MIT.
• Custom licences. A licence named after the model (for example "Kimi K3 License" or "Muse Model License") sets its own terms. These can include limits that standard licences do not have. Meta's Llama 3 licence is a well-known example: if a licensee's products had more than 700 million monthly active users in the month before the model's release, the Llama 3 licence says they "must request a license from Meta, which Meta may grant to you in its sole discretion".

Before using an open-weight model in a product, read its licence on the vendor's repository page; each model page here links the vendor's own source.

Which AI models have open weights?

Many of the language models in our AI models directory publish their weights. A selection, with the licence each vendor states (positions are read live from LMArena's text leaderboard):

Model	Developer	Licence	LMArena text
Kimi K3	Moonshot AI	Kimi K3 License	1487 Elo (rank 18, max effort)
Qwen3.8-Max	Alibaba	Licence named "qwen3.8-max"	1482 Elo (rank 24, max effort)
GLM-5.3	Z.ai (Zhipu AI)	GLM-5.3 License (custom)	1479 Elo (rank 28, max effort)
GLM-5.3-Flash	Z.ai (Zhipu AI)	MIT	1474 Elo (rank 41)
DeepSeek V4-Pro	DeepSeek	MIT	1458 Elo (rank 64)
DeepSeek V4.1-Flash	DeepSeek	MIT	1475 Elo (rank 38, max effort)
MiMo V2.6 Pro	Xiaomi	MIT	1479 Elo (rank 27)
Muse Spark Open	Meta	Muse Model License	no LMArena rating
Nemotron 3.5 Lightning	NVIDIA	OpenMDW 1.1	1351 Elo (rank 216)

The full list of open-weight models, with each one's licence, is in the self-hosting answer on the AI model rankings page.

Where do you download open-weight models?

Most are published on Hugging Face, a site that hosts model files, and many vendors also link them from their own documentation. A model's repository page usually has the weight files, the licence, a model card describing what the model is for and how it was evaluated, and example code. See how to read an AI model card.

To actually run a downloaded model, tools such as Ollama, llama.cpp and vLLM load the weights and serve them; see how to run AI models locally.

How much hardware does an open-weight model need?

Roughly, the model's size in parameters times the bytes stored per parameter, plus working memory. Weights stored at 16-bit precision take 2 bytes per parameter, so a 70-billion-parameter model needs about 140 GB just for its weights. Compressed ("quantised") to 4 bits, the same model needs about 35 GB. Large mixture-of-experts models need memory for their total parameter count, even though only a fraction is used per token; see mixture of experts explained.

This is why the largest open-weight models are usually run on servers with several GPUs, while smaller ones run on a single graphics card or a recent laptop.

Why do vendors publish model weights?

Vendors give their own reasons on release, and they differ; the practical effects for users are the same. Publishing the weights lets people:

• run the model without sending data to the vendor, which matters for private or regulated data;
• run it offline or on their own infrastructure, at their own cost;
• fine-tune it for a specific task or language;
• study how it behaves, which researchers cannot do with an API-only model.

A closed model keeps those options with the vendor. Many companies publish some models' weights and keep others closed.

Words to know

• Weights (parameters): the numbers a model learns in training; the model file is mostly these.
• Open-weight model: a model whose weights are published for download.
• Closed (API-only) model: a model used only through the vendor's app or API.
• Open-source AI: under the OSI definition, a system whose weights, full training and running code, and detailed training-data information are all available under open terms.
• Licence: the legal terms for using, changing and sharing the model.
• Fine-tuning: training a published model further on your own data.
• Quantisation: storing weights with fewer bits so the model takes less memory.
• Hugging Face: the site where most open-weight models are published.

Related guides

• How to run AI models locally
• Mixture of experts explained
• How to read an AI model card
• The open-source software behind AIblogly