Listen to this lesson
Unlock audio and more
Audio streaming, downloadable PDFs and certificates come with Plus and Pro.
What it means
The distinction matters more than the industry's vocabulary suggests. Open weights means the trained parameters are downloadable, so you can run the model on your own hardware. Open source, properly used, additionally means a license granting broad rights to use, modify and redistribute — typically with the training code and data pipeline available too.
Most models described in the press as "open source" are open-weight releases under custom licenses. Those licenses routinely carry restrictions a genuinely open license would not: revenue or user thresholds above which terms change, mandatory attribution, bans on training competing models, and in some cases geographic exclusions barring use in particular countries entirely.
A downloadable model, a public repository and a permissive-sounding announcement are all compatible with a license that forbids the thing you want to do.
Why it matters
Teams adopt open-weight models specifically to gain control — over cost, over data residency, over their own roadmap. A license restriction discovered after deployment defeats that entirely, and these terms are load-bearing for exactly the commercial use cases that motivate self-hosting in the first place.
What people get wrong
That "open source" in a headline describes the license. It usually describes downloadability. The phrase has become standard press shorthand for any release you can download, and it gets applied to licenses that a software-licensing body would not recognize as open source at all. A public repository, downloadable weights and a permissive-sounding announcement are all compatible with terms that forbid the exact thing you want to do.
That the restrictions are theoretical. They bite precisely at the commercial uses that motivate self-hosting in the first place. Recent releases have carried revenue and monthly-active-user thresholds above which a separate written agreement is required; clauses barring use of the outputs to train a competing model; mandatory attribution; and — the one most likely to invalidate a deployment outright — geographic carve-outs excluding entire jurisdictions, in at least one case naming the United States and the EU. A team that read "open" as "ours to run" can discover after deployment that it never was.
That a license carries forward to the next generation. Terms are set per release. A lab that published permissively before can tighten when a model gets good enough to be worth protecting — which is exactly when you most want to use it. Never infer this generation's terms from the last one's.
In practice
Read the actual license file, not the announcement or the coverage. Four questions settle most cases: is commercial use permitted at all; are there revenue or user thresholds; is attribution or a no-compete clause mandatory; and is your country excluded? Note that restrictions often apply to the downloadable weights while the hosted API stays generally available, so say which path is actually open to you.
Where this shows up
Tools and models in our catalog.
Llama 4Meta's open-weight frontier model family. Maverick: MoE 400B/17B active, 1M context, 1,417 Elo. Scout: 10M context. Most-downloaded open-weight models. Community license (free under 1M MAU).
DeepSeek R1First open-source reasoning model matching OpenAI o1. MIT license. R1-0528 adds JSON output and function-calling. Distilled variants 1.5B-70B. Banned on government devices in several countries, and ordered off US federal civilian systems in October 2026.
Mistral Large 3Mistral's previous flagship open-weight model: 675 billion-parameter MoE (41 billion active), 256K context, multimodal, modified MIT license. Succeeded in October 2026 by the 1 trillion-parameter Mistral Large 4, which ships under a different license; Large 3 remains the smaller option to self-host.
Qwen3.8-MaxAlibaba's flagship from August until Qwen4-Max replaced it on November 16, 2026: a 2.4 trillion-parameter mixture-of-experts model with about 95 billion active per request and roughly a one-million-token context. Its weights are downloadable under a custom Qwen license, which makes it the largest Qwen model you can self-host.
Kimi K3Moonshot AI's flagship — a 2.8 trillion parameter mixture-of-experts model activating 16 of 896 experts per token, with a 1 million token context window and native vision. Weights published in July 2026 under a custom Kimi K3 License. Kimi K3.1 Pro, a post-trained version released October 23, 2026 with cyber-related training data removed, runs on Moonshot's apps and API, and its weights have been public since November 4, 2026 under the same commercial terms as K3.