Match open base model to its license consideration
Drag each answer to line up with its matching prompt
Llama 3 / 3.1 / 3.3
Apache 2.0 for most sizes (with a small number of variants under their own commercial license): fine-tuned derivatives generally inherit the permissive terms.
Mistral 7B / Mixtral
MIT license for the weights, with model card specific commercial use clarifications: among the most permissive frontier quality open releases.
Qwen 2 / 2.5
Apache 2.0 for the open weight base models: broadly permissive commercial use including for fine-tuning and redistribution.
DeepSeek-R1 / V3-0324
Meta Llama Community License: generally permissive but has a 700M monthly active users clause: companies above that user threshold need a separate license from Meta.
Gemma 2 / 3
Google's own Gemma Prohibited Use Policy: permissive for most commercial use but with explicit category exclusions you must review.
Open base model licenses fall into three buckets: Apache 2.0, MIT, and custom vendor licenses that add clauses like Meta's user threshold or Google's use-case exclusions.
Imagine borrowing tools from five workshops to build something you'll sell. Some workshops hand over the tool and say 'use it however you like, just keep our name on the box': that's Apache and MIT. One workshop says 'free for almost everyone, but if your shop gets enormous, come back and sign a deal': that's Meta's Llama rule about huge user counts. Another says 'free, but here's a list of things you're not allowed to build': that's Google's Gemma policy. The tool works the same in every case. The difference is the paperwork attached, and that paperwork decides whether your business can ship the thing you built.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Fine-tuning is not a clean-room operation. The moment you load an open base model and update its weights, your result is a derivative work. The base model's license follows you into production. This is why license consideration belongs at the start of a fine-tuning project, alongside the benchmark and cost analysis, not bolted on at legal review.
The trap is intuitive but wrong: weights that download for free feel like public-domain assets. They are not. These five model families all publish open weights anyone can pull from Hugging Face, yet they sit under three different legal regimes. Two carry standard open-source licenses with almost no conditions. One is even more minimal. The remaining two are custom vendor licenses that read like a contract, with clauses that can quietly disqualify a specific deployment.
The three license buckets
The cleanest mental model splits open base models into three buckets, and every family here lands in exactly one.
The first bucket is standard permissive open source. Mistral 7B, Mixtral, and most Qwen 2 / 2.5 sizes ship under Apache 2.0. This is an OSI-approved license: you may use, modify, fine-tune, and redistribute commercially. The two obligations are keeping the license notice and respecting an explicit patent grant. DeepSeek-R1 and V3-0324 sit nearby under MIT, which grants the same freedoms in even fewer words and without the patent clause. The original DeepSeek-V2 and V3 weights used the custom DeepSeek Model License, so check the exact checkpoint.
The second and third buckets are custom vendor licenses. Meta's Llama Community License is permissive for the vast majority of users but adds a scale-based gate. Google's Gemma combines its own terms with a Prohibited Use Policy that gates by use case.
Apache and MIT are open source in the formal sense. The Llama and Gemma terms are vendor contracts that happen to allow most commercial use. The download mechanics look identical, so the difference is invisible until you read the text.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
| Model family | License | Key condition to check |
|---|---|---|
| Llama 3 / 3.1 / 3.3 | Meta Llama Community License | Separate Meta license required above 700M monthly active users |
| Mistral 7B / Mixtral | Apache 2.0 | Broad commercial use; keep the notice and attribution |
| Qwen 2 / 2.5 | Apache 2.0 (most sizes) | A few variants carry their own license; check per size |
| DeepSeek-R1 / V3-0324 | MIT (weights) | Model-card commercial clarifications; among the most permissive |
| Gemma 2 / 3 | Gemma terms + Prohibited Use Policy | Categorical use-case exclusions you must review |
Real products, models, and research that use this idea.
- Hosted fine-tuning platforms such as Together.ai and Fireworks offer Llama and Mistral bases. The customer stays bound by each base model's license, including the Meta user-count clause, when serving the result.
- DeepSeek V4 Preview shipped its open weights under MIT, so a team fine-tuning it needs less license review than with a vendor license.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does a base model's license propagate to a fine-tuned derivative you redistribute?
Treat the fine-tune as a derivative work. The base terms carry forward, so Apache requires notice retention and Llama carries its threshold and naming clauses into your release.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Assuming every open-weight model is open source. Several use custom vendor licenses with conditions that can block your specific commercial use, even when the weights download freely.
60 second bullets to scan on the way to the call.
Which families are Apache 2.0 versus MIT versus custom vendor
What the Llama 700M monthly active users clause actually triggers
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.