Open Weights vs Open Source AI: What the Labels Actually Mean

“Open source AI” has become one of those phrases that everybody uses and almost nobody defines. A model can be free to download, free to run commercially, and still fail every traditional test of open source. Another can be published by a lab that calls it open while attaching conditions that would make a lawyer wince.

If you are choosing a model to run on your own hardware, the labels matter more than usual, because they determine whether you can actually use the thing for work.

Open Weights Means You Get The Numbers

An open-weights model is one where the trained parameters are published for download. You can fetch the file, run it on your own machine, and generate text without calling anybody’s API.

That is a genuinely large concession from a lab. Weights are the expensive part; training them costs a fortune in compute. Publishing them means anybody can run the model forever, offline, with no per-token fee and no possibility of the vendor watching the traffic.

What open weights does not tell you is anything about the training data, the training code, the data-cleaning pipeline, or the reinforcement learning recipe. You get the finished cake, not the ingredients list. You cannot reproduce the model, and you cannot audit what went into it.

For most professional use, that is an acceptable gap. You care that the model runs locally and produces good text, not that you can rebuild it from scratch.

Open Source Means Something Narrower

The Open Source Initiative has maintained a definition of open source software for more than two decades, and it requires more than a downloadable artifact: source code, freedom to modify, freedom to redistribute, and no discrimination against fields of use.

Applying that to models is awkward, because a model’s “source” is arguably its training data, which almost no lab publishes. The OSI eventually produced a separate Open Source AI Definition to address exactly this mismatch. Under it, a model needs sufficient information about training data for a skilled person to recreate a substantially equivalent system.

Very few well-known models clear that bar. Most of what the press calls open source AI is more accurately open weights.

The Licenses Are Where The Real Constraints Live

This is the part that actually affects your work, and it varies more than people expect.

Some open-weights models ship under Apache 2.0 or the MIT license. These are genuinely permissive: commercial use, modification, and redistribution are all fine, with an attribution requirement and no user-count ceiling. As of mid-2026 this camp includes OpenAI’s gpt-oss models, Google’s Gemma 4 family, and most of the major Chinese releases including DeepSeek, Qwen, GLM, and Kimi.

Others ship under a bespoke community license. Meta’s Llama family is the best-known example: broadly usable, but with an acceptable-use policy and a clause that requires a separate license above a very large monthly-active-user threshold. Fine for a company of forty people, relevant if you are a company of forty million.

And some models are published under research-only terms, which forbid commercial use entirely. This trap is easy to fall into, because a research-licensed model downloads exactly like a commercial one. Wrivio hit this directly: an earlier build shipped a model under a non-commercial research license, and swapping to Apache 2.0 alternatives was not optional once we noticed.

What To Check Before You Commit To A Model

Four questions, in order of how often they bite people:

Can I use the output commercially? Read the license text, not the blog post announcing it. Research licenses and non-commercial clauses are the single most common surprise.

Does the license survive redistribution? If you plan to bundle the model with software, you need distribution rights, not just usage rights.

Are there use-case restrictions? Acceptable-use policies attached to community licenses can exclude categories that matter to you.

Can the terms change under me? Weights you have already downloaded cannot be recalled, which is one of the underrated advantages of local deployment. An API’s terms can change next quarter; a file on your disk cannot.

Why This Matters For Writing Tools Specifically

If you rewrite client emails, patient notes, or contract language, the reason to care about open weights is not ideology. It is that an open-weights model can run entirely on your machine, which means the text never becomes someone else’s log entry.

That is a categorically different privacy position from “the vendor promises not to train on your data.” A promise is a policy, and policies change. A model running on your CPU with the network unplugged is an architecture, and architectures do not have quarterly reviews.

Wrivio’s Local mode exists for this reason: it runs an Apache 2.0 open-weights model in-process on your PC, so a rewrite of a confidential paragraph involves no network call at all. If you want the reasoning behind picking small models for this job, we wrote about why small models often beat big ones for rewriting.

Common Questions

Is an open-weights model less capable than a closed one?

At the frontier, closed models still lead on the hardest reasoning and coding tasks. For constrained tasks like rewriting, summarizing, and reformatting, the gap in 2026 is small enough that most people cannot pick the winner blind. The open-weights field has closed most of the distance on ordinary work.

Can I fine-tune an open-weights model?

Usually yes, if the license permits it. Apache 2.0 and MIT models are unambiguously fine-tunable. Community licenses often require you to carry naming conventions or license terms into the derivative.

Does open weights mean free?

Free to download and run, yes. Not free to operate: you pay in hardware, electricity, and your own time. The economics still favor local for high-volume small tasks, which is most professional writing.

Where do I find the license for a model?

The model card on Hugging Face states the license, and serious labs link the full text. If a model card is vague about licensing, treat that as a finding rather than an oversight.

Download Wrivio for Windows to run an Apache 2.0 open-weights model behind a hotkey, with your text never leaving your PC.