What are open-weight AI models? Open weights vs open source

When an AI company says it has released an "open" model, it usually means one specific thing: you can download the weights. That is a big deal, because it lets you run the model on your own computer or server, keep your data private and tweak the model for your own tasks. But "open weights" is not the same as "open source", and the difference matters if you plan to build on it.

This explainer covers what weights are, how open-weight models differ from closed ones and from true open-source AI, which licence terms to look for and what hardware you need to run one yourself.

What are model weights?

A large language model is, at its core, a huge set of numbers called parameters or weights. During training, these numbers are adjusted billions of times until the model gets good at predicting the next piece of text. The finished file of weights is the model: together with a relatively small piece of code that describes its architecture, it is everything needed to generate answers.

Model sizes are quoted in parameters: a "7B" model has about 7 billion, a "70B" model about 70 billion. Big frontier models run to hundreds of billions or even trillions; many now use a mixture-of-experts design in which only part of the network (the "active" parameters) is used for each token, which makes them faster to run but not smaller to store.

Open weights vs closed models

With a closed model such as the ones behind ChatGPT, Claude or Gemini, the weights never leave the company's servers. You send a prompt through an app or an API, and you get text back. You can't inspect the model, run it offline or change how it was trained.

With an open-weight model, the developer publishes the weights, usually on a hub such as Hugging Face, so anyone can download them. Well-known open-weight families include Meta's Llama, Mistral's models, Alibaba's Qwen, DeepSeek, Google's Gemma, OpenAI's gpt-oss and Moonshot's Kimi; the newest example is Kimi K3.1 Pro, whose weights Moonshot has promised by 4 November.

Why open weights are not the same as open source

In software, "open source" means you get the source code and the freedom to use, study, modify and share it. For AI, the Open Source Initiative published an Open Source AI Definition in 2024. It says a truly open model should come with the weights, the code used to train and run it, and enough information about the training data that a skilled person could build a substantially equivalent system.

Most open-weight releases don't meet that bar. Companies typically publish the final weights and inference code, but not the training data, which may include licensed or scraped material, and not the full training pipeline. You can use and fine-tune the model, but you can't fully reproduce or audit how it was made. That is why many researchers prefer the term "open-weight" for these releases.

Licences: the fine print that matters

"Downloadable" does not mean "do whatever you like". Licences fall roughly into three groups:

• Permissive licences such as Apache 2.0 or MIT allow commercial use, modification and redistribution with few conditions. Several Qwen, Mistral, DeepSeek and gpt-oss models use them.
• Custom community licences allow broad use but add rules. Meta's Llama licence, for example, requires very large companies (above a monthly-user threshold) to get a separate licence, and includes an acceptable-use policy.
• Research or non-commercial licences let you experiment but not build a paid product.

Before you ship anything, read the licence for the exact model version, check whether it restricts certain uses or users, and note any attribution requirements.

Why people choose open-weight models

• Privacy: prompts and documents never leave your machine or your company's servers.
• Offline use: a local model works on a plane or in a secure network with no internet.
• Cost control: after the hardware, there's no per-token bill; at scale, self-hosting can be cheaper than an API.
• Customisation: you can fine-tune on your own data, or use small adapters (LoRA) to teach a style or domain.
• Transparency and research: researchers can probe how models behave, test safety measures and compare results.

There are trade-offs too. The very best models are often still closed, you are responsible for updates and security, and safety guardrails can be removed by anyone who fine-tunes the weights. That last point is why some developers trim risky abilities before release; Moonshot, for instance, says it removed cyber-attack training data from Kimi K3.1 Pro.

What hardware do you need to run one?

The rule of thumb is about memory. Each parameter takes 2 bytes at 16-bit precision. Quantisation shrinks this: at 8 bits it's about 1 byte per parameter, and at 4 bits about half a byte, with modest quality loss for most chat tasks. Add some extra for the context window.

GPU memory (VRAM) or Apple-style unified memory is what counts most; system RAM works but is slower. Our guides to choosing a laptop for local AI and how much RAM you need go deeper, and NPUs can help with small on-device models.

How to try an open-weight model

1. Pick a tool. Free apps such as Ollama, LM Studio and llama.cpp download models and run them with a chat interface or a local API.
2. Start small. A 4-bit 7B or 8B model runs on most 16 GB laptops and is good for drafting, summarising and coding help.
3. Check the format. Local tools commonly use quantised GGUF files; Python frameworks use safetensors.
4. Read the model card. It lists the licence, intended use, languages and known limitations.
5. Scale up only if needed. Bigger models are smarter but slower and need far more memory.