Open weights, explained—the AI models you can download and keep

Last week Meta released an AI model with 816 billion parameters, and anyone could download it. A day later a startup called Reflection released one with 501 billion. Meanwhile a model that reads photos and documents now fits in about 3GB on a phone, and an image generator runs on an 8GB graphics card.

None of these require an account with the company that made them. You download a file, and the model is yours to run, change and build on.

This is what people mean by open weights, and it is one of the most useful ideas in AI to understand. It explains why there is an "Open Weights" section in every AI This Week, how you can train a model on your own writing at your desk, and why the AI you use doesn't have to live on someone else's servers.

What a weight is

An AI model is, underneath, a very large list of numbers. During training, a computer reads an enormous amount of text (or images, or audio) and keeps adjusting those numbers until the model gets good at predicting what comes next. Those numbers are the weights, also called parameters. The Open Source Initiative describes them as the learned values that sit on top of a model's design and turn an input into an output.

When someone says a model has 816 billion parameters, they mean the list holds 816 billion of those numbers. Training is the expensive part: months of work on thousands of specialized chips. Once it's done, the weights are just a file, or a set of files.

A company that keeps its weights private lets you use the model only through its app or its paid connection for programs (an API). A company that publishes its weights hands you the file itself.

Open weights isn't quite open source

With ordinary software, "open source" means you get the source code: the human-readable instructions you can study, change and rebuild.

Weights are different. You can run them and adjust them, but you can't read them the way you read code, and you can't rebuild them without the training data and the training code. Think of a bakery that gives you a finished loaf, free to eat, slice, toast and even sell. That is open weights. Open source would also hand over the recipe, the list of where the flour came from and the oven settings.

In October 2024 the Open Source Initiative published a definition of open source AI that asks for all three: the weights, the full code used to train and run the model, and enough information about the training data that a skilled person could build a substantially equivalent system. Few big releases meet that bar. Most of what gets called "open" in AI, including the models above, is open weights.

That still gives you a lot. You can run the model on your own hardware, keep your data on it, fine-tune it for your own work and keep using it if the company changes its mind.

Read the license

"Open" covers a range of terms, and the license says which one you're getting. Two from the past week show the spread:

• Reflection's Beam uses Apache 2.0, a standard open-source license: use it, change it, sell what you build, no conditions about who you are.
• Meta's Muse Spark Open uses Meta's own Muse Model License. Use, fine-tuning and distillation are allowed. A company with more than $20 million in revenue that offers the model as a hosted service has to sign a separate agreement with Meta.

For a person or a small business, both are free to use. The differences matter when you build something big on top of one.

Why companies give them away

Training a large model costs a fortune, so giving one away seems odd. The companies have reasons.

Their business is somewhere else. In a 2024 letter, Mark Zuckerberg wrote that selling access to AI models isn't Meta's business model. Meta makes money from its apps, so a free model that becomes a standard costs it little and helps it a lot.

A standard attracts builders. The same letter compares open models to Linux, the free operating system that outgrew its paid rivals once enough people built on it. When developers write tools around your model, everyone improves it for you.

Nobody wants to rent their foundation. Zuckerberg also wrote that Meta doesn't want to depend on a competitor's closed platform. Developers feel the same way about a model that a vendor can change, reprice or switch off.

Some do sell it, too. Reflection publishes Beam's weights and also runs it on its own paid service and through cloud partners such as AWS, Cloudflare and Google Cloud. Many people would rather pay someone to run a 501-billion-parameter model than buy the hardware.

There is a competitive reason as well. Some of the strongest open models in recent years have come from Chinese labs such as Alibaba's Qwen team and DeepSeek, and U.S. companies now describe their releases in those terms. TechCrunch's headline on Beam's debut called it a model to rival Chinese ones, and Meta calls Muse Spark Open "the most capable U.S. open-weight model to date."

What you actually download

The weights are only one of three pieces.

The weights are the file: billions of numbers and nothing else. On their own they do nothing, the way a music file doesn't play itself.

The runner is a program that loads the weights into memory and does the math that turns your question into an answer. Free runners such as llama.cpp, MLX and vLLM do this work, and most open models are published in formats they can read.

The chat window is how you talk to it: a text box, a conversation history, a place to drop in a document. It sends your words to the runner and shows you what comes back.

Desktop apps such as LM Studio and Ollama bundle all three. They find and download the weights, run them, and give you a chat window, which is why they are the easiest way to start.

Making a model versus using one

Training and using a model are very different jobs.

Training sets the numbers. A computer reads a huge amount of text and adjusts billions of weights over and over, for weeks or months, across thousands of specialized chips. It happens once, and it is where the cost goes.

Using a model only reads the numbers. Your question is broken into tokens, words or pieces of words, and run through the weights to pick the most likely next token. That token is added to the text and the process repeats, one token at a time, until the answer is finished. Nothing in the weights changes.

That is why a model that took a data center to make can answer questions on a laptop. Your computer isn't redoing the training, only following the result. Whether it can, and how fast, depends mostly on how much memory it has and how quickly it can read from it.

What fits on your computer

Start with a small model. An 8-billion-parameter model has 8 billion numbers to keep in memory.

A model's weights are numbers, and storing those numbers takes memory. Quantization stores them with less precision, a little like rounding a measurement from 12.783 to 12.8.

"4-bit" describes how much space each weight uses. Compared with 16-bit weights, that's roughly a quarter of the storage for the weights themselves. The computer needs additional memory to run the model and handle your conversation.

The model keeps the same number of parameters, but their values become less precise. That can affect answer quality; how much depends on the model and how it was compressed.

So the 8-billion-parameter model takes about 16GB at 16 bits and about 4GB at 4 bits. Add some room for the runner and your conversation, and it fits comfortably on a laptop with 16GB of memory. OpenAI's gpt-oss-20b, released in August 2025 under Apache 2.0, is built to run in 16GB too. Models of that size write, summarize, answer questions about your documents and help with code.

Not every parameter works at once. Many new models are mixture of experts: they hold many specialized sub-networks and use only a few for each token. That makes them faster, though all the weights still have to be stored somewhere. MiniCPM-V 4.7 keeps its experts in storage and loads them as needed, which is how a 35-billion-parameter model runs in about 3GB on a phone.

Then there are the giants. Muse Spark Open uses 42 billion of its 816 billion parameters per token. Its full weights are about 1.5 terabytes. In 4-bit form it needs about 493GB, and Apple ran it across four Mac Studios linked by Thunderbolt, producing 36.7 tokens a second. The arithmetic is the same as for the 8-billion-parameter model, just a hundred times bigger.

What you can do with one

Keep your data at home, for good. A model on your own computer never sends your medical notes, client files or diary anywhere, and it works on a plane. It doesn't get retired, repriced or quietly changed either: the file you have today works the same next year.

Make it yours. Because you have the weights, you can train a small add-on, called an adapter, that teaches the model a style or a task from your own examples. The Workshop guide to LoRA walks through exactly that, using the open-weights model behind Nate's own essay-voice adapter.

Build on it. A business can put an open model inside its product without paying per question, and keep it running on the same version for as long as it likes.

Where to start

You can try an open-weights model this afternoon without typing a command.

1. Install LM Studio or Ollama. Both are free and work on Mac, Windows and Linux.
2. Pick a model that fits your memory. On a laptop with 16GB, start with something around 8 to 20 billion parameters in a 4-bit build. The apps show which ones fit.
3. Give it a real job. Paste in a long email thread and ask for the decisions and deadlines. Ask it to rewrite a paragraph more plainly. Then turn off your Wi-Fi and do it again.

If you want to look around first, Hugging Face is where most open models are published, each with a page listing its size, license and what it's good at. When a new one lands in AI This Week, that page is the place to check whether it will run on your machine.