These four model families can all be downloaded and run on your own servers, but their licences are not the same. For a firm, the licence and the way you plan to run the model matter more than any league table. We compare licence terms only, and we do not rank models.
Last reviewed: 17 October 2026. Facts checked. We opened the Llama 3.1 licence text on GitHub (Meta's llama-models repository), the Qwen3-8B page, the DeepSeek-V3 page and its model licence file, and the Mistral-7B-Instruct-v0.3 page, all on Hugging Face. We did not check newer releases in each family. We give no benchmark scores, parameter counts (beyond one stated below) or release dates.
A disclosure. CORAA runs open-source models on India-based infrastructure. This post does not rank models, and it compares licences, not quality.
What "open-source LLM" means here
An LLM is a program that predicts text. "Open-source" is often used loosely for any model whose files you can download. Strictly, it means the licence lets you use, change and share it. Some models are better called open-weight, because the terms add conditions. Our companion post on open-weight models and confidentiality explains the difference.
The licences, side by side
| Family | What we read | Licence as stated | Points to note |
|---|---|---|---|
| Llama (3.1 text) | Licence text on GitHub | Meta's community licence | Needs a "Built with Llama" notice if you distribute. Model names built from Llama outputs must begin with "Llama". A separate licence is needed above 700 million monthly active users. Use must follow Meta's Acceptable Use Policy |
| Qwen (Qwen3-8B page) | Model page | Apache-2.0 | The page gives 8.2B parameters. Other Qwen models may differ |
| Mistral (7B-Instruct-v0.3 page) | Model page | Apache-2.0 | The model card says the model has no moderation mechanisms |
| DeepSeek (V3 page) | Model page and licence file | Code under MIT. Weights under a Model License | The page says commercial use is supported. The licence has use-based restrictions, including on fully automated decisions that adversely affect legal rights |
These are the terms for the specific versions we opened. Each family has many models. Open the page for the exact one you plan to run.
What the licence means for a CA firm
Most audit firms will run a model internally and will not redistribute it. That makes some clauses less relevant. A few still matter.
- Use-based restrictions. Both the Llama and DeepSeek licences point to use rules. Read them against your planned use. Do not assume an audit task is safe because it sounds routine.
- Redistribution. If you build a product on a model and ship it to clients, naming and notice terms can apply.
- Version drift. A licence can change between versions of the same family.
- Your own lawyer. This post is not legal advice.
Fit for audit work
Licence is the first filter. Fit is the second. Fit means whether the model does your task well enough, on your documents, with your checks.
- Reading and summarising long documents. Test on your own anonymised contracts and minutes.
- Extracting figures from messy PDFs. Test with known answers and count misses.
- Drafting working-paper text. Check it against the source, line by line.
- Language. If clients send Hindi or regional-language material, test that specifically.
We do not claim any family wins at these. Public benchmarks rarely match audit work. Our task-by-task model guide explains how to run a fair comparison.
Running it: three practical questions
- Where will it run? Your server, a hosting partner or a vendor API change the confidentiality picture. See open-source vs self-hosted vs API key.
- What will it cost? Hardware, staff time and monitoring all count. Our cost and ROI guide walks through it.
- How will you keep it stable? Pin the exact version, as we cover in the post on model change control.
A selection routine
- List your three real tasks.
- Shortlist two or three models by licence.
- Run each on the same fifty anonymised files.
- Score accuracy, misses and reviewer time.
- Record the model, version and licence in the engagement file.
The AI audit tool evaluation checklist gives a structure for that record.
Frequently Asked Questions
Which open-source LLM is best for audit?
We do not rank models. The right one depends on your task, your data and your hosting. Test two or three on your own files.
Is Llama free for commercial use?
The Llama 3.1 licence we read grants use under conditions, including a notice requirement and a special licence above 700 million monthly active users. Read the licence for the exact version you use.
Is Mistral open source?
The Mistral-7B-Instruct-v0.3 page lists Apache-2.0. Other Mistral models may have different terms.
Can we use a model with a Chinese developer?
We did not find or verify any official restriction for this post, so we say nothing about legality. Check your clients' requirements, your regulator's guidance and the licence text, and choose where it runs.
Does the licence cover the outputs?
Licences differ on outputs. Read the relevant section for the version you use.
Related Articles
Statutory facts on this page are checked against their sources, and the page says where it relied on secondary reporting. How we verify · Report an error