Key takeaways
- Open weights means the trained parameters are published. Open source, as the Open Source Initiative defines it for AI, also needs the training code, detailed data information, and terms that allow any use without asking permission.
- Kimi K3.1 Pro, whose weights were published on Hugging Face in early November 2026, is described by Moonshot AI as 'open-weight'. Its licence is free to use, but a Model as a Service business above US$20 million revenue in 12 months needs a separate agreement first.
- The Llama 4 Community License adds a 700 million monthly active user test, 'Built with Llama' attribution, a 'Llama' name prefix for derived models, and an Acceptable Use Policy that withholds multimodal rights from people and companies based in the EU (end users of products built on those models are not affected).
- Apache 2.0 has no size thresholds or use policy, includes a patent licence, and is listed by the OSI as an approved licence. Its duties are mainly about keeping notices and marking changes.
- Before you ship, read three things in the licence: who it excludes, what you must display, and what triggers a separate deal.
On this page
- Open weights vs open source vs closed: at a glance
- The four parts of an AI model release
- What "open source" means, according to the OSI
- Kimi K3.1 Pro vs Llama 4 vs Apache 2.0: the licence terms compared
- Reading the Kimi K3.1 Pro licence
- Reading the Llama 4 Community License
- Reading Apache 2.0
- Worked example: three startups, one model
- Which should you choose?
- A five-question licence check before you ship
Open weights vs open source comes down to what is released and on what terms. An open-weight model publishes its trained parameters for download, under any licence the maker chooses. Open source, as the Open Source Initiative (OSI) defines it for AI, goes further: it also needs the code used to train and run the model, detailed information about the training data, and a licence that lets anyone use, study, modify and share the system for any purpose. Most downloadable models today are open weights. The licence file is what tells you whether you can ship one.
This page explains the difference in plain terms, then compares three real licences side by side: the new Kimi K3.1 Pro licence from Moonshot AI, Meta's Llama 4 Community License, and Apache 2.0. Every licence term below is quoted from the licence text itself. This is a plain-English reading, not legal advice; for a real product decision, have a lawyer read the licence you depend on.
Open weights vs open source vs closed: at a glance
| Closed model | Open weights | Open source AI (OSI) | |
|---|---|---|---|
| Weights downloadable | No, API only | Yes | Yes |
| Training code | No | Usually not | Required |
| Training data information | No | Usually limited | Required in detail |
| Licence | Terms of service | Anything the maker chooses | Any use, no permission needed |
| Run it on your own servers | No | Yes, within the licence | Yes |
| Rebuild it from scratch | No | No | In principle, yes |
The four parts of an AI model release
A trained model is not one thing. When a company "releases" a model, it can publish any mix of four parts:
Weights
The billions of numbers the model learned during training. With these you can run the model and fine-tune it.
Code
The software to run the model (inference) and, separately, the software that trained it. Inference code is common. Training code is rare.
Data
The text, images or other material the model learned from, or at least a detailed description of it. This is the part most often held back.
Licence
The legal terms that say who may use the other three, for what, and with which duties attached.
"Open weights" says something about the first part only. "Open source" is a claim about all four, measured against a published standard.
What "open source" means, according to the OSI
For ordinary software, the reference is the OSI's Open Source Definition. It has ten criteria. Two matter most for AI licences:
- Criterion 5, No Discrimination Against Persons or Groups: "The license must not discriminate against any person or group of persons."
- Criterion 6, No Discrimination Against Fields of Endeavor: the licence "must not restrict anyone from making use of the program in a specific field of endeavor".
Criterion 2 adds that source code "must be the preferred form in which a programmer would modify the program." For a model, weights alone are hard to call the preferred form, because you cannot retrain or meaningfully audit the model without the code and data behind it.
That is why the OSI wrote a separate Open Source AI Definition (version 1.0). It asks for four freedoms: to use the system "for any purpose and without having to ask for permission", to study it, to modify it, and to share it. It then lists what the "preferred form" for modifying a machine-learning system must include:
- Data Information: "Sufficiently detailed information about the data used to train the system so that a skilled person" could build a substantially equivalent system. This means a complete description of the training data and a listing of publicly available data and where to get it. It does not demand the full dataset itself.
- Code: "The complete source code used to train and run the system," under OSI-approved licences.
- Parameters: "The model parameters, such as weights or other configuration settings," under OSI-approved terms.
So a model can be open weights and still fall short of open source in two separate ways: by withholding training code or data information, or by a licence that excludes some users or uses.
Kimi K3.1 Pro vs Llama 4 vs Apache 2.0: the licence terms compared
Kimi K3.1 Pro arrived on Hugging Face in early November 2026. Its model card lists 2.8T total and 104B activated parameters and a context of 1,048,576 tokens, and it calls the model "an open-weight, natively multimodal agentic model." Note the wording: open-weight, not open source. The card says it is "released under the Kimi K3.1 Pro License" and does not say anything about releasing training data or training code.
The table compares its licence with the Llama 4 Community License Agreement (Llama 4 Version Effective Date: 5 April 2025, the newest Llama licence on Meta's site when we checked) and the Apache License, Version 2.0 (January 2004).
| Question | Kimi K3.1-Pro License | Llama 4 Community License | Apache 2.0 |
|---|---|---|---|
| Basic grant | Free of charge, "to deal in the Software without restriction", including to sell and fine-tune | "non-exclusive, worldwide, non-transferable and royalty-free limited license" | "perpetual, worldwide" copyright licence, no charge |
| Size or revenue trigger | Model as a Service business with group revenue over US$20 million in any 12 consecutive months needs a separate agreement | Over 700 million monthly active users on the Llama 4 release date must request a licence from Meta | None |
| Attribution in your product | Display "Kimi K3.1-Pro" if over 100 million monthly active users or over US$20 million monthly revenue | Display "Built with Llama" | Keep notices and the NOTICE file contents; no product branding |
| Naming derived models | No rule | Must start with "Llama" | No rule; no trademark rights granted |
| Use restrictions | No acceptable use policy in the licence | Acceptable Use Policy incorporated by reference | None |
| Region limits | None | Multimodal rights not granted to EU-domiciled people or EU-based companies (end users of products are not affected) | None |
| OSI position | No OSI statement found | OSI says Llama 2 and Llama 3.x licences are not Open Source | Listed by OSI as an approved licence |
Reading the Kimi K3.1 Pro licence
The LICENSE file is titled "Kimi K3.1-Pro License" (with a hyphen, while the model card writes "Kimi K3.1 Pro") and carries "Copyright (c) 2026 Moonshot AI". It opens like the MIT licence: "Permission is hereby granted, free of charge, to any person (the "Licensee") obtaining a copy of this software". The main duty for everyone is to keep "The above copyright notice and this permission notice" in "all copies or substantial portions of the Software."
Then come five short numbered sections. Two add real conditions.
The Model as a Service clause
The licence defines "Model as a Service" as "giving a third party access to language model inference or fine-tuning (e.g., via API) in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data." In plain terms: selling raw access to the model.
If you run that kind of business and "the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months", then "the Licensee must enter into a separate agreement with Moonshot AI" "before using the Software or its derivative works for any commercial purpose."
Two things are carved out: "(a) end-user products with model capabilities solely embedded within specific features or harnesses," and "(b) mere relaying of requests to models hosted by others." A writing app with a built-in assistant is not a Model as a Service business under this wording. An inference API that lets customers send any prompt and set parameters is.
The display clause
Commercial products or services "that have more than 100 million monthly active users, or more than 20 million US dollars" "(or equivalent in other currencies) in monthly revenue," must show the name: "Kimi K3.1-Pro" must be "prominently displayed on the user interface of such product or service."
The exemptions
Neither condition applies to "internal use of the Software, defined as any use that does not make the Software, its outputs, or its underlying capabilities available to third parties", nor to "any use of the Software accessed through Moonshot AI's official products or certified inference partners." The licence has no acceptable use policy and no rule against using outputs to train other models. It ends with an "as is" warranty disclaimer.
Reading the Llama 4 Community License
Meta's licence is longer and adds duties at smaller scale than Kimi's:
- Attribution for anyone who distributes or makes it available, or offers a product that contains it. Section 1.b.i requires anyone who distributes or makes available the Llama Materials, or a product or service that contains them, to display "Built with Llama" on related materials. If you use Llama materials or outputs to create or improve an AI model that you distribute or make available, its name must begin with "Llama".
- A user threshold. Section 2 applies if, "on the Llama 4 version release date," your products or services had "greater than 700 million monthly active users in the preceding calendar month." Those companies must request a licence, which Meta may grant at its sole discretion.
- A use policy. Section 1.b.iv requires compliance with the "Acceptable Use Policy for the Llama Materials." The Llama 4 Acceptable Use Policy lists prohibited areas such as "Violence or terrorism" and "Guns and illegal weapons".
- An EU carve-out for multimodal models. The policy says that "With respect to any multimodal models included in Llama 4, the rights granted under Section 1(a)" are not granted to individuals domiciled in the EU or companies with their principal place of business there. End users of products built on those models are not covered by the exclusion.
These are the same kinds of terms the OSI objected to. In 2023 it said the Llama 2 licence "puts restrictions on commercial use for some users" and "restricts the use of the model and software for certain purposes (the Acceptable Use Policy)", against criteria 5 and 6. In February 2025 it wrote that "Llama 3.x is still not Open Source by any stretch of the imagination." That post predates Llama 4 and does not assess the Llama 4 licence.
Reading Apache 2.0
Apache 2.0 was written for software. It grants a "perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable" patent licence on top of the copyright licence. That patent licence ends for you if "You institute patent litigation against any entity" over the work.
The duties in Section 4 apply when you redistribute:
- "You must give any other recipients of the Work or Derivative Works a copy of this License".
- "You must cause any modified files to carry prominent notices stating that You changed the files".
- Keep "all copyright, patent, trademark, and attribution notices from the Source form of the Work".
- If there is a "NOTICE" file, include "a readable copy of the attribution notices contained within such NOTICE file".
Section 6 says the licence "does not grant permission to use the trade names, trademarks, service marks, or product names" of the licensor. There is no user threshold, no revenue threshold and no use policy. The OSI's licence page marks it as an "Open Source Initiative Approved License." A model under Apache 2.0 still is not open source AI by the OSI's definition unless its code and data information are also released.
Worked example: three startups, one model
Take three companies thinking about Kimi K3.1 Pro, and read each against the licence text quoted above.
- A legal-research SaaS runs the model on its own servers to summarise documents inside its app. Customers never send raw prompts to the model. This looks like an end-user product "with model capabilities solely embedded within specific features", so the Model as a Service clause does not bite. The display clause starts only above 100 million monthly active users or US$20 million in monthly revenue. Its main duty is keeping the licence notice.
- An inference provider sells API access to the model, letting customers set prompts and parameters. That fits the Model as a Service definition. If the company and its affiliates earned more than US$20 million in total over any 12 consecutive months, it needs a separate agreement with Moonshot AI before commercial use.
- A hospital's internal team uses the model only for staff, with nothing exposed to customers. That is "internal use" as the licence defines it, so both scale conditions are exempt.
Swap in Llama 4 and the picture changes for the first company: a product that contains Llama must show "Built with Llama", any fine-tune it distributes or makes available must have a name starting with "Llama", and it must follow the Acceptable Use Policy. If the company is based in the EU and the model is multimodal, the policy withholds the licence rights. Swap in an Apache 2.0 model and all three companies only need to handle notices and change markings.
If you decide not to self-host and call a hosted model through an API instead, the licence terms above mostly stop being your concern, and keeping your API key secret becomes the main risk. Our guide on what to do about a leaked API key covers that side.
Which should you choose?
The right model depends on how you will use it. These cards apply the licence terms quoted above to common situations.
Internal tools only
The Kimi K3.1-Pro licence exempts internal use from its Model as a Service and display conditions, defined as use that does not make the model, its outputs or its capabilities available to third parties. Apache 2.0 has no such conditions either.
An app with an embedded assistant
Kimi's Model as a Service clause carves out end-user products with model capabilities embedded within specific features. With Llama 4, expect "Built with Llama" and the Acceptable Use Policy. Apache 2.0 asks only for notices and change markings.
Selling API access to the model
Under Kimi's licence this is Model as a Service, so a separate agreement is needed above the revenue trigger. Apache 2.0 has no revenue trigger.
An EU company using multimodal features
The Llama 4 Acceptable Use Policy withholds the multimodal rights from EU-domiciled people and EU-based companies. End users of a product built on those models are not affected. The Kimi licence and Apache 2.0 have no region limits.
You need the "open source" label
The OSI's Open Source AI Definition is the reference most people mean. It needs the code and data information released as well, and terms that allow any use. Apache 2.0 is an approved licence, but a model under it is not open source AI unless those parts are released too.
You want the simplest terms
Apache 2.0 has no user or revenue thresholds and no use policy. Its duties are notices and marking changes.
A five-question licence check before you ship
- What exactly is released? Weights only, or also training code and data information? This decides whether you can audit or rebuild it.
- Who is excluded? Look for user counts, revenue numbers, regions and named competitors.
- What triggers a separate deal? Find the clause and check whether your business model, not just your size, fits it.
- What must you display or keep? Notices, a "built with" line, a model name in the interface, or a naming rule for fine-tunes.
- Is there a use policy? If yes, read it in full, because it is part of the licence.
The labels on a model page ("open", "open-weight", "open source") are marketing until you match them to these answers. For more explainers like this one, see the Dexaloop Tech hub.
Questions
What is the difference between open weights and open source?
Open weights means the trained parameters are published for download, under whatever licence the maker picks. Open source, under the OSI's Open Source AI Definition, also requires the complete training and inference code, detailed information about the training data, and terms that let anyone use, study, modify and share the system for any purpose.
Is Llama open source?
Meta publishes the Llama weights, but the OSI said in 2023 that the Llama 2 licence is not Open Source and in February 2025 that 'Llama 3.x is still not Open Source'. Its 2023 post gave the reasons as the commercial limit for some users and the Acceptable Use Policy. In 2025 it also pointed to the exclusion of people in the European Union. The Llama 4 licence still contains a 700 million monthly active user test and an Acceptable Use Policy.
Can I use Kimi K3.1 Pro commercially?
The licence grants free use, including the right to sell, as long as you keep the copyright and permission notice. Two conditions apply at scale: a Model as a Service business whose group revenue passes US$20 million over any 12 consecutive months must sign a separate agreement with Moonshot AI first, and products with more than 100 million monthly active users or more than US$20 million in monthly revenue must display 'Kimi K3.1-Pro' in the interface.
Does open source AI require releasing the training data?
Not in full. The OSI's definition requires 'Data Information': a complete description of the training data, a listing of publicly available training data and where to get it, and enough detail that a skilled person could build a substantially equivalent system. It also requires the complete source code used to train and run the system.
Is Apache 2.0 better than a custom model licence?
It is simpler and better known. Apache 2.0 has no user or revenue thresholds, no use policy, and an explicit patent licence, and the OSI lists it as an approved licence. A custom licence can still be generous, as Kimi's is for most companies, but you have to read its specific triggers.
Sources
- Kimi K3.1 Pro model card (Moonshot AI, Hugging Face) (accessed 9 November 2026)
- Kimi K3.1 Pro repository metadata (Hugging Face API) (accessed 9 November 2026)
- Kimi K3.1-Pro License, LICENSE file (Moonshot AI, Hugging Face) (accessed 9 November 2026)
- Apache License, Version 2.0 (Apache Software Foundation) (accessed 9 November 2026)
- Apache License 2.0 (Open Source Initiative licence page) (accessed 9 November 2026)
- Llama 4 Community License Agreement (Meta) (accessed 9 November 2026)
- Llama 4 Acceptable Use Policy (Meta) (accessed 9 November 2026)
- The Open Source Definition (Open Source Initiative) (accessed 9 November 2026)
- The Open Source AI Definition 1.0 (Open Source Initiative) (accessed 9 November 2026)
- Meta's LLaMa license is not Open Source (OSI blog, 20 July 2023) (accessed 9 November 2026)
- Meta's LLaMa license is still not Open Source (OSI blog, 18 February 2025) (accessed 9 November 2026)
Written from the sources above and checked against them before publishing. Found a mistake? Email support@dexaloop.com and we will correct it.