AI 101 — What are open-weight models?

An open-weight model is an AI model whose trained parameters — the billions of numbers that encode everything it has learned — are released for anyone to download, run, modify, and build on. The weights are the model: once they are public, the technology itself is out in the world, not locked behind a company's servers. That single distinction divides AI into two camps — the closed, API-only frontier and the open ecosystem — and it is the most consequential licensing question in the field right now.

Why it's in the news

This week handed the open camp its biggest wins in months. DeepSeek released the open weights for V4 Pro — build 0813, its production flagship — on Hugging Face under an MIT license: roughly 890 GB of weights, a 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters and a 1-million-token context window, downloadable by anyone. The timing was strategic: the weights landed the same week DeepSeek raised its API prices by up to 470 percent, giving builders the frontier model free and charging only those who want it served. Days earlier, Alibaba's Qwen team open-sourced Qwen3.8-2.4T-A95B, its first Max-class model. And the policy world is catching up: Wired reports the White House is "almost certain" to extend its AI oversight framework to open models once they reach frontier capability — a framework that today reviews only closed models, and one that becomes nearly unenforceable the moment weights are public. Even OpenAI, which built its early brand on keeping models closed, now publishes a case for open weights as a way to spread AI to everyone.

The mental model

Think of a model as a cake and its training as the baking. The training data and the training code are the recipe; the weights are the finished cake. A closed model means you can only eat the cake in the restaurant — you pay per slice, through an API. Open weights mean the restaurant hands you the cake to take home: you can slice it, frost it, study its texture, even bake small variations by adjusting it. But you still don't have the recipe — the training data and code stay in the kitchen. "Open source" AI goes further: it publishes the recipe too, so anyone could, in principle, bake the cake from scratch. Most licensing arguments in AI today are arguments about how much of the kitchen to open.

An everyday analogy

A famous chef publishes her signature dish in three ways. Closed: you can only order it at her restaurant. Open weights: she puts the finished dish in a takeout box with a full ingredient list — you can reheat it, garnish it, weigh every component, and tweak the seasoning, but you can't know the exact cooking technique or where the ingredients came from. Open source: she writes the complete cookbook entry — every step, every pan, the sourcing notes, even the failed experiments. That is the difference between running a model, fine-tuning a model, and being able to rebuild one. Most people are perfectly happy with the takeout box — which is why open weights, not full open source, is where the industry's energy has gone.

Common misconceptions

"Open weights are the same as open source." No. The Open Source Initiative's official definition of open-source AI requires access to the training data and the code that built the model — not just the finished weights. Meta calls its Llama models open source, but they ship under a custom license that requires Meta's permission above 700 million monthly users; the OSI does not consider them open source. DeepSeek's MIT license on V4 Pro is permissive and genuine, but it covers the weights, not the training data, which remains undisclosed.

"Open weights mean no rules." The license still sets the terms — MIT is about as free as licensing gets, but many open-weight models ship with restrictions on commercial use, redistribution, or what you can build with them. And governments are moving in: the White House's expected expansion of oversight to open models would add a regulatory layer on top of licensing, even though pre-release review is practically impossible once weights are downloadable.

"Anyone can run them on a laptop." An 890 GB model with 1.6 trillion parameters is not a phone app. It needs serious hardware — though the ecosystem responds fast: within hours of the V4 Pro release, unsloth had published quantized versions that shrink the model dramatically while keeping most of its capability, which is how open models reach people with modest machines.

"Open means transparent." The weights are billions of numbers; no human can "read" them. Openness is about access — you can run the model, study its behavior, and fine-tune it — not interpretability. You know the dish's exact composition, not why it tastes the way it does.

"Open models trail closed ones." That gap keeps shrinking. DeepSeek claims its 0813 build jumped from 72.1 to 87.9 percent on Terminal-Bench 2.1 — though every figure so far is vendor-reported, and independent runs have not landed yet. The honest summary: open-weight frontier models now trade blows with closed flagships, and the closed labs' argument that openness necessarily costs capability is getting harder to defend.

Where to learn more

The Open Source Initiative's Open Source AI Definition is the definitive statement of what "open source" means for AI — and why most open-weight releases don't qualify. The Hugging Face page for DeepSeek V4 Pro is a good example of what a real release looks like, including the license terms. For the closed side of the argument, OpenAI's "Open Weights and AI for All" explains why even an API-first company came around to releasing weights.

Related reading: If the 1-million-token window in this story piqued your interest, our What is a context window? explainer covers why that number matters. And if you're wondering why closed labs guard their models' inner workings so closely, What is chain of thought? explains the reasoning the weights produce — and the fight over keeping it private.

If anyone can download a frontier model's weights, does pre-release government review still make sense — or is the cat permanently out of the bag? Tell us in the comments.

Sources: DeepSeek-V4-Pro-0813 on Hugging Face · OSI — The Open Source AI Definition · OpenAI — Open Weights and AI for All · Wired — The White House is going to expand its AI policy · AI Midday — DeepSeek ships V4 Pro, ending its flagship's four-month preview · unsloth — DeepSeek V4 Pro GGUF (Hugging Face)