Open model licences compared: what you can actually ship

Sort models by what triggers a condition, not by the word open. A kind of product, such as a coding assistant, can need a separate licence at any revenue. Every revenue limit we compare starts at 20 million US dollars. Check what your product is first, then what it earns.

Can we ship this open-weight model in our product, and which licence clauses decide it? Open-weight means the trained parameters are published, so you can run the model yourself. We are Lyceum, a sovereign European AI inference provider for open-source models. Our serverless models run in Paris and Finland, and prompts and outputs are processed, not stored, and never used for training.

This article reads the licence files of nine models on our models page, as they stood on 19 October 2026. One hypothetical startup runs through them. It sells a stand-alone coding assistant and earns US$5 million a year. It has 200,000 monthly active users, meaning people who use the product in a month. The startup and its numbers are assumed, not a real company.

Open weights and open source are different grants

Open weights means the weights are published. Weights are the set of learned parameters that turn an input into an output. Open source is a stricter grant. The Open Source AI Definition requires the freedom to use a system for any purpose and without asking permission. It also requires a description of the training data and the code used to derive the parameters.

A licence that requires a separate licence at a revenue limit fails the freedom to use a system without asking permission, because you must ask. Such a model is open weights, not open source, and your approval should say so.

The question comes up more often now. Since January 2026, the most capable open-weight models have trailed frontier closed models by an average of four months on Epoch AI's capability index. More products ship on them, and each needs its licence read.

Where do you read that licence? On Hugging Face, the licence name sits in the model's card, the metadata at the top of its README.md. A custom licence text sits in a LICENSE file in the repository. Our model pages name the licence as well, and the GLM-5.3 page links the file. Read the file, not the name.

Licences of the open models we serve, compared

Three of the nine models ask only that you keep a notice. The other six add a display duty, a revenue limit, a kind of business, or several of these. A notice is the copyright and licence text you keep with copies of the software. A display duty means showing the model's name on your product.

The licensee is the person or company that holds the licence. Revenue limits count the licensee and its affiliates, which are related companies such as a parent or sister company. The table shows what each licence file asks of the licensee.

Licence files read on 19 October 2026. Model as a Service and AI Work Assistant are defined in the clauses section below.

Revenue and user thresholds that change your terms

A threshold is a limit above which a condition starts, and these licences use three kinds. Display duties use monthly active users or monthly revenue. Separate-licence rules use total revenue over a 12-month stretch, and MiniMax M3 uses yearly revenue.

Check each figure against the limit as a share of it:

monthly revenue = yearly revenue / 12 = 5,000,000 / 12 = 416,667
share of limit  = your figure / the licence limit
display check   = 416,667 / 20,000,000 = 2.1%   and   200,000 / 100,000,000 = 0.2%

The hypothetical startup sits at 2.1% of the monthly revenue limit and 0.2% of the user limit. No display duty applies on Kimi K2.7 Code, Kimi K3 or the two Qwen models. The table shows the other conditions for the same startup.

The example favours the startup in one way. Most limits sit far above US$5 million, which makes the revenue test look easy. The business-type rule on Qwen3.8 Flash Next ignores revenue, and MiniMax M3 applies from the first commercial use. The example does not show whether your product counts as a Model as a Service or AI Work Assistant. We assumed it is an AI Work Assistant and not a Model as a Service, and the licensor decides.

Model as a Service clauses and internal use

Model as a Service means giving a third party access to model inference or fine-tuning, for example via an API. The third party must get meaningful control over inputs, parameters or training data. The GLM-5.3 licence excludes end-user products with model capabilities solely embedded within specific features, and the mere relaying of requests to models hosted by others. The Kimi K3 definition uses the same words. The Qwen definition excludes only the mere relaying of requests.

An AI Work Assistant is an independent AI-powered product primarily designed for AI-assisted coding or office productivity, according to the Qwen licences. The licence excludes an assistant that is a feature of a product whose main purpose is something else. A stand-alone coding assistant is an independent product built for coding, so the startup triggers the Qwen3.8 Flash Next rule. A coding helper inside an accounting product would not.

Internal use is the exception to the Model as a Service and AI Work Assistant rules. Kimi K3 defines it as any use that does not make the software, its outputs or its underlying capabilities available to third parties. The Qwen licences carry the same exemption from the separate-licence rule. Your customers receive the outputs of a shipped product, so a product feature is not internal use. A script that only your own engineers run may be.

Display duties, API use and routing aliases

A display duty is a condition about your interface, and the licences word it differently. Kimi K3, Kimi K2.7 Code and the Qwen licences ask for the model's name on the user interface of the product. MiniMax M3 asks for "Built with MiniMax M3" on a related website, user interface, blogpost, about page or product documentation.

MiniMax M3 also names APIs. Its Commercial Use definition covers products offered for a fee that use the software. It also covers the commercial use of APIs provided by or for the software. Its appendix bans military use. We report the text and do not say how it applies to calls through any particular API. Ask MiniMax or counsel.

A routing alias is a model name under which a provider picks the model for each request. Lyceum Router is one: we choose the model for each request and bill it as that model. Its model page says it routes to MiniMax M3 today. MiniMax M3 is the only model here with a non-commercial grant, so a Router call can land under the strictest terms. Its model page shows the MiniMax Community License.

The word today matters, because the model behind an alias can change and the licence behind it with it. Pin a model ID instead. A pinned ID always reaches the same model, so its licence file stays the one you read:

from openai import OpenAI

client = OpenAI(api_key="YOUR_LYCEUM_API_KEY", base_url="https://api.lyceum.technology/openai/v1")
response = client.chat.completions.create(
    model="qwen/qwen3.8-27b",  # pinned: always Qwen3.8 27B under Apache 2.0
    messages=[{"role": "user", "content": "Hello!"}],
)

A licence check you can repeat per model

A repeatable check turns the licence into one record you keep next to the pinned model ID. Fill in one record per model before it ships, and again when you switch models. This is the record for the Apache 2.0 model from the example:

model_id: qwen/qwen3.8-27b
licence: Apache 2.0
licence_file: https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/LICENSE
read_on: 2026-10-19
display_duty: no
revenue_limit: none
business_type_trigger: none
internal_use_exception: not needed
confirmed_by: <name of the person who read the file>

If any line says yes, names a trigger or is unknown, ask the licensor or counsel before you ship. The record for Qwen3.8 Flash Next would read `business_type_trigger: AI Work Assistant`. That line blocks the launch until you hold a separate licence.

This article describes licence text. It is not legal advice, and it does not say whether a customer of an API counts as the licensee. For a wider supplier review, see our AI vendor risk assessment checklist.

Pick the model on our models page, open its licence file and pin that model ID in your code.