The best open source LLM is the one whose licence fits how you plan to use it, not just the one with the best-known name. Makers use the words differently: some publish open weights only, some publish weights plus code and reports, and some release models under a permissive licence such as Apache 2.0. Read the licence line before the family name.
Search for an open source LLM and you'll find makers calling their models open, open-weight, fully open or open-source. Those words aren't interchangeable, and it's the licence line, not the label, that tells you what the maker is releasing.
This list covers ten model families, using only what each maker's own page says, read on 2 Nov 2026 (as of Nov 2026). It's alphabetical and in no particular order. There are no benchmark scores, parameter counts or rankings here: just who makes each family and how the maker describes it.
Open model families and their makers
| Tool | Type | Good for |
|---|---|---|
| AI21 Jamba | Hybrid Transformer-Mamba language models | Weights under the Jamba Open Model License |
| Arcee AI | Open models from an American research lab | The Trinity open-weight model family |
| Falcon from TII | Open-source LLM series from TII | Falcon-H1 and Falcon-E models |
| Gemma | Google DeepMind's open models | Versions for mobile, IoT devices and personal computers |
| IBM Granite | Apache 2.0 licensed models from IBM | On-prem deployment and enterprise workloads |
| Mistral AI | Frontier and specialist open-weight models | Tailoring models to your organisation |
| Nous Research | Open AI training and research lab | The Hermes Agent harness |
| Olmo from Ai2 | Fully open language model family | Weights, code and reports for each variant |
| OpenAI gpt-oss | Open-weight reasoning models | Two models under the Apache 2.0 license |
| Zyphra | Open models and inference | Customising models in the cloud or your own environment |
Ten open model families, one line each
AI21 Jamba
Best for: Reading the Jamba-1.5 paper, whose page says the model weights are publicly available under the Jamba Open Model License.
Arcee AI
Best for: Following Trinity, which the lab describes as an open-weight model family that runs anywhere.
Falcon from TII
Best for: Exploring Falcon-H1 and Falcon-E, from what the Technology Innovation Institute calls its open-source large language model series.
Gemma
Best for: Building with Google DeepMind's family of open models, with versions pitched at mobile and IoT devices and at personal computers.
IBM Granite
Best for: Downloading models released under an Apache 2.0 license, which IBM positions for secure, on-prem AI deployment and enterprise workloads.
Mistral AI
Best for: Starting with frontier and specialist open-weight models and making them specific to your organisation.
Nous Research
Best for: Following a lab that calls itself a pioneer in open AI training and research, and that created the Hermes Agent harness.
Olmo from Ai2
Best for: Picking an Olmo 3 variant and exploring its weights, code and reports, which Ai2 calls a fully open language model and complete model flow.
OpenAI gpt-oss
Best for: Running the two gpt-oss open-weight reasoning models, which OpenAI releases under the Apache 2.0 license.
Zyphra
Best for: Using open-source frontier models you can customise and consume in the cloud or in your sovereign environment, in the maker's words.
Open weights, fully open or a named licence
Here's the split a searcher wouldn't guess: 'open source' means something different on each maker's page. Mistral's page title promises 'frontier open models you own', and it describes its range as 'open-weight models'. Arcee describes Trinity as an 'open-weight model family'. Ai2 goes further, calling Olmo 'our fully open language model and complete model flow', with weights, code and reports for each variant.
Then there's the licence itself. IBM says its Granite models are 'Apache 2.0 licensed', and OpenAI says gpt-oss is available under the 'flexible Apache 2.0 license'. AI21 says the Jamba weights are 'publicly available under the Jamba Open Model License', a licence named after the model. TII calls Falcon an 'open-source large language model series', and Zyphra's page promises 'open-source frontier intelligence'.
- Open weights: makers whose pages say open-weight models or an open-weight model family.
- Fully open: weights plus code and reports, in one maker's words.
- Named licences: Apache 2.0 on two pages, and a licence named after the model on another.
Independent labs, big companies and research institutes
Who makes a model is the other way to sort this list. Some are independent labs: Arcee says it 'develops open-weight models in the U.S. and releases them for everyone', and Nous Research describes itself as 'a pioneer in open AI training and research'. Its site header carries the words 'Open Source' and 'MIT License'.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
Others come from big companies. Google DeepMind lists the Gemma 4 family under open models, IBM says Granite is 'powering secure, on-prem AI deployment', and OpenAI's announcement describes gpt-oss as 'open-weight reasoning models'. Then there are research outfits: the Technology Innovation Institute says its Falcon models reflect a commitment to 'making advanced AI accessible and impactful', and Ai2 publishes a technical report beside its Olmo models.
Where these open models can run
Several makers talk about where a model runs, too. Arcee says Trinity 'runs anywhere'. Mistral says you can make its models 'specific to your organization'. IBM's Granite page is built around on-prem deployment, and one maker on this list says you can customise its models and use them in the cloud or in your own sovereign environment.
If you plan to run a model on your own hardware, that's the line to look for on the maker's page, alongside the licence. It tells you whether the maker expects you to host the model yourself or reach it through someone's cloud.
Before you download a model
- What exact licence does the maker name, and where is it published?
- Does the maker release weights only, or weights plus code and reports?
- Is the licence a standard one, such as Apache 2.0, or one named after the model?
- Will I run the model on my own hardware, in a cloud, or both?
- Do I need to customise the model for my organisation?
- Who'll read the licence terms before the model goes into a product?
Do you make an open source LLM or open-weight model? Tool makers can ask to be considered for a featured place on this list.
My advice before you pick an open model
Read the licence line before you fall for a family name. If you need code and training details as well as weights, look for a maker that says fully open. If you'd like a standard licence your legal team already knows, look for one that names it, such as Apache 2.0. And if a maker uses its own licence, read that licence in full before a model goes anywhere near a product.
New to all this? My explainer on what a large language model is covers the basics, and if you'd like to try a model on your own machine, here's how to use Ollama.
Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.
Frequently asked questions
What's the difference between open source and open weights?
On these makers' own pages, open weights means the model weights are released, as in open-weight models or an open-weight model family. One maker goes further and calls its model fully open, publishing weights, code and reports for each variant. Others name a licence, such as Apache 2.0.
Which of these models use the Apache 2.0 licence?
Two makers on this list name it: one says its models are Apache 2.0 licensed, and another says its two open-weight reasoning models are available under the Apache 2.0 license. Check the licence file for the exact model you download.
Which model on this list has its own licence?
One maker says its model weights are publicly available under a licence named after the model, the Jamba Open Model License. Read that licence on the maker's own site for what it allows before you build on it.
Which of these makers are independent labs?
Several describe themselves as labs on their own pages: one calls itself an American research lab building open intelligence, one calls itself a pioneer in open AI training and research, and one says it makes frontier open models you own.
Can I run these models on my own hardware?
Some makers say so in their own words. One lists secure, on-prem AI deployment, one describes an open-weight family that runs anywhere, and one says you can use its models in the cloud or in your sovereign environment. Check each model's docs for what it needs.
Where do I find the licence for each model?
On the maker's own model page or repository. This list quotes each maker's licence wording as it appears on that maker's page, read on 2 Nov 2026, so check the current terms before you use a model in anything you ship.
