TECHTechExplainer NO. ---

Open-Weight vs Open-Source AI: What's the Difference, and Why It Matters to You

Open-weight and open-source AI are not the same thing. Learn the difference, how to read a model license, and what it takes to run one yourself.

By Rowan Ashby Published 20.10.26 08:00 UTC 5 min read

Code on computer monitor (Unsplash) — Markus Spiske markusspiske / Wikimedia Commons / CC0 · Source

By Rowan Ashby · October 20, 2026

Almost every week a company announces an "open" AI model. In the week of October 13, 2026, alone, Meta released Muse Spark Open and DeepSeek updated its V4.1 Flash model. Neither charges for its weights, yet the two come with very different rules: DeepSeek's can be downloaded straight away, while Meta's require a request form and manual approval on Hugging Face, plus a custom license. "Open" does a lot of work in AI headlines, and it does not always mean what people assume.

This guide explains the difference between open-weight and open-source AI, shows how to check a model's license in a few minutes, and gives a rough way to tell whether a model will run on your own hardware.

The short answer

Open-weight means you can obtain the model's trained parameters (its "weights") and run it yourself. What you are allowed to do with it depends on the license.
Open-source AI, in the stricter sense used by the Open Source Initiative (OSI), means you also get the code to train and run the model and detailed information about its training data, on terms that let anyone use, study, change and redistribute it.

Every open-source model is open-weight. Not every open-weight model is open source.

What are "weights"?

An AI model is, at its core, a very large set of numbers that training adjusts until the model gets good at predicting text, images or code. That finished set of numbers is the weights, or parameters. A model with "816 billion parameters" has a weights file holding roughly that many numbers.

With the weights and suitable software, you can run a model on your own machine without sending your data to its maker. That is the practical benefit of any open-weight release: you decide where it runs, who sees your prompts, and whether to keep an older version. What weights alone do not give you is the recipe: without training code and data information, you cannot fully see how the model was built or reproduce it.

What "open source" means for AI

The OSI, which maintains the widely used Open Source Definition for software, announced version 1.0 of its Open Source AI Definition on October 28, 2024. In short, it requires that anyone be free to run the system for any purpose without asking permission, to examine how it works, to change it (including what it outputs), and to pass it on to others, modified or not.

For those freedoms to be practical, the definition requires three things: enough detail about the training data, all of the source code needed to train and run the system, and the weights themselves. A model whose weights are downloadable but whose training data is undocumented does not meet that bar, whatever the press release says.

The four tiers you will actually see

Tier	What you get	Example
Closed / API-only	Access through an app or paid API; no weights	Most frontier chat models from large labs
Open-weight, custom license	Weights under the company's own terms, often with limits	Meta's Muse Spark Open (Muse Model License)
Open-weight, standard permissive license	Weights under a well-known license such as MIT or Apache 2.0	DeepSeek-V4.1-Flash-1017 (MIT)
Open source (OSI sense)	Weights, training code and data information, under open terms	Pythia (EleutherAI), OLMo (AI2), Amber and CrystalCoder (LLM360), and T5 (Google), which passed the validation phase of the OSI's definition process

On that last row, the OSI says its validation results "are not certifications of any kind," and that it reviews legal documents rather than individual AI systems.

The middle two tiers are where most people get caught out.

Custom licenses can carry real conditions. Muse Spark Open's license requires products to display "Built with Muse Spark Open." A company offering the model to others as an API-style service, with combined gross revenue above US$20 million in any 12-month period, must notify Meta and sign a separate agreement first. Meta's usage policy also states the model is not intended for people under 18. For a hobbyist that matters little; for a startup it matters a lot.

Standard permissive licenses are simpler. MIT and Apache 2.0 allow commercial use, modification and redistribution, mainly asking you to keep the copyright and license notices. Apache 2.0 adds an explicit patent license from contributors and asks you to mark files you changed. But a permissive license on the weights says nothing about whether the training data is documented.

How to check a model's license in five minutes

Most open-weight models are published on Hugging Face. Before building on one:

Read the license tag on the model page. "mit" or "apache-2.0" means a standard license; a custom name means you need to read the full text.
Open the LICENSE file in the "Files" tab. The tag is a label; the file is what applies.
Search the license for revenue or user thresholds ("revenue," "monthly active users," "service"). These clauses most often affect businesses.
Look for attribution rules that require a notice in your product or documentation.
Check for a separate usage policy and any access form. These can add conditions on top of the license itself.

If you are building a commercial product, have the license reviewed by someone qualified. This guide is general information, not legal advice.

Can you run it yourself? A rough rule of thumb

"Downloadable" is not the same as "runs on your laptop." A quick estimate:

At 16-bit precision (BF16), each parameter takes 2 bytes, so an 8-billion-parameter model needs about 16 GB for the weights alone.
Compressed ("quantized") to 4 bits, each parameter takes about half a byte, so the same model needs roughly 4–5 GB, within reach of many recent laptops and gaming GPUs.
An 816-billion-parameter model like Muse Spark Open needs about 1.6 TB at 16-bit and roughly 400 GB or more at 4-bit: data center territory, far beyond a typical personal computer.

These figures cover the weights only; real use needs extra memory for the text the model is working on, so treat them as a floor.

Why it matters even if you never download a model

Open-weight releases put price pressure on paid APIs, let researchers test claims independently, and let organizations keep sensitive data in-house. Licenses decide who gets those benefits. When you next see "open" in an AI headline, ask three questions: Can I get the weights? What does the license let me do? Can I see how it was trained?