What it is
OLMo is a family of language models made by the Allen Institute for AI, a nonprofit research institute in Seattle founded in 2014 by the late Microsoft co-founder Paul Allen. The institute calls itself Ai2 (it used to write AI2), and since the Olmo 3 release in late 2025 it has written the model's name as "Olmo". The letters stand for Open Language Model.
Like Llama, Qwen, or DeepSeek, OLMo is a model you can download and run yourself. What sets it apart is how much else comes with it. Most companies that share a model publish only its weights, the billions of learned numbers explained in Parameters. Ai2 also publishes the training data, the code used to clean that data and train the model, the settings, the training logs, the tests it used, and hundreds of in-between versions of the model saved along the way. It calls this "fully open".
A useful comparison is a cake. An open-weights model is a finished cake you are allowed to take home, slice, and decorate. A fully open model is the cake plus the recipe, a list of every ingredient with where it was bought, the oven temperature, and photos taken every ten minutes while it baked. You probably won't bake it again yourself, because the oven here costs millions, but you can check what went into it and why it turned out the way it did.
The generations at a glance
Product details change quickly, so treat this as a snapshot at the time of writing and check Ai2's site for the latest.
- OLMo 1 (1 February 2024). A 7-billion and a 1-billion-parameter model, trained on at least 2 trillion tokens from Dolma, Ai2's own open text collection. Ai2 released the weights, training code, training data, more than 500 checkpoints per model, training logs, and evaluation code, all under the Apache 2.0 licence.
- OLMo 2 (26 November 2024, with a 32B model in March 2025). Models of 7, 13, and 32 billion parameters, with instruction-following versions trained using Tülu 3, Ai2's openly published recipe for the finishing stages of training.
- OLMoE (2024 to 2025). A small mixture-of-experts model that Ai2 put in a free iPhone app in February 2025.
- Olmo 3 (20 November 2025) and Olmo 3.1 (12 December 2025). Models at 7 and 32 billion parameters, including "Think" versions that write out their reasoning before answering, trained on a new data collection called Dolma 3.
- Olmo Hybrid (5 March 2026). An experimental 7-billion-parameter model that mixes the usual transformer design with a different kind of layer, which Ai2 says reached Olmo 3's score on one standard test using 49 percent less training data.
What "fully open" means, and how others compare
Words like "open" and "open source" get used loosely in AI, so it helps to separate three levels.
- Closed models, such as GPT, Claude, and Gemini, run on the company's computers. You use them through an app or an API and never see the weights.
- Open-weights models let you download the weights. Licences vary. Meta's Llama licence adds conditions on who may use it and how. Many Qwen models, and some Mistral models, use Apache 2.0, a standard permissive licence, and DeepSeek released R1 under the MIT licence. These are generous terms. But none of these companies publishes the full training data, so outsiders can't check what the model learned from or rebuild it.
- Fully open models add the data, code, and training record. OLMo is the best-known family in this group. Others include Pythia from the nonprofit EleutherAI and the models from the LLM360 project.
In October 2024 the Open Source Initiative, which maintains the long-standing definition of open-source software, published its Open Source AI Definition 1.0. It says an open-source AI system must share the code used to train and run it, the parameters, and enough information about the training data that a skilled person could build a substantially equivalent system. Notably, it does not require the data itself to be published, a compromise some critics dislike. During the testing of that definition, the initiative's volunteers found that OLMo, Pythia, LLM360's Amber and CrystalCoder, and Google's T5 passed, while Llama 2, Grok, Phi-2, and Mistral's Mixtral did not. That was an assessment of specific older versions, not a permanent label on each company.
Why it matters
It makes AI something you can inspect, not just use. With most models, questions like "was this book in the training data?" or "did the model learn this answer or just memorise it?" can only be answered by the company, if at all. With OLMo, anyone can search the data. That matters for researchers, but also for journalists, authors wondering whether their work was used, and regulators trying to check claims.
Researchers can run real experiments. Science needs repeatable results. If you want to know how a model picks up a skill during pretraining, you need the in-between versions and the exact data they saw. Ai2's checkpoints let university teams without a large budget study questions that used to be answerable only inside a big lab, such as when memorisation happens, or how changing the data mix changes the result. See Datasets for why the data shapes so much of a model's behaviour.
Auditors can check claims instead of trusting them. When a company says its model scores well on a test, an outsider can't easily tell whether the test questions were in the training data, which would make the score look better than it should. This problem, called contamination, is covered in Benchmarks. With open data it can be checked directly.
Educators get a complete worked example. Courses on how language models are built can point students at a real, modern model where every step is documented, from raw web pages to a finished chat assistant.
It shows who can afford openness. Ai2 is a nonprofit, not a company selling subscriptions. In August 2025 the US National Science Foundation and NVIDIA announced $152 million over five years for an Ai2-led project to build fully open AI models for science. In March 2026, when Ai2's chief executive Ali Farhadi stepped down, the chair of its board told the news site GeekWire that "the cost to do extreme-scale open model research is extraordinary". Openness at the very largest scale is expensive, which is one reason fully open models remain smaller than the leading commercial ones.
Real-world examples
The first fully open release (February 2024)
On 1 February 2024, Ai2 released OLMo 7B together with a paper titled "OLMo: Accelerating the Science of Language Models". The paper argued that the most capable models had become closed off behind commercial interfaces, which made it hard to study their biases, risks, and abilities scientifically. The release was unusually complete. Alongside the weights came the Dolma dataset and the tools to rebuild it, training code, over 500 saved checkpoints for each base model, the full training logs, and the evaluation code. Ai2 also credited partners that supplied computing power and help, including AMD, Finland's CSC IT Center for Science, and Databricks.
The licensing also changed in the open. Dolma had first been released in August 2023 under Ai2's own ImpACT licence, which had conditions. On 15 April 2024 Ai2 moved it to ODC-BY, an open data licence that only asks for credit, saying researchers were using the data in ways it had not expected and needed more freedom. The models themselves have used Apache 2.0 from the start.
OLMoTrace: checking where an answer came from (April 2025)
On 9 April 2025, Ai2 launched OLMoTrace in its free online playground. After the model answers, you can press a button that highlights long, unusual phrases in the answer that appear word for word in the training data, and shows the documents they appear in. It works by searching trillions of words of training text in seconds, using an index technique Ai2 calls infini-gram. No company that keeps its data private can offer this.
Ai2's own launch post gave two telling examples. In one, its 13-billion-parameter model wrongly said its knowledge ended in August 2023, when its pretraining data actually ended in December 2022. OLMoTrace pointed to the likely cause: examples in the later, smaller training stage that mentioned that date. Ai2 said it then removed knowledge-cutoff statements from the training data of its 32-billion-parameter model. In the other, a correct answer to a hard competition maths problem turned out to contain a calculation that appeared several times in the training data, so the model could have memorised it rather than worked it out. It is the kind of evidence that the Interpretability and Hallucination pages describe as hard to get.
One caution: a match shows that the same words appear in the data, not that the model definitely learned them from that document.
A reasoning model you can trace (November 2025)
On 20 November 2025, Ai2 released Olmo 3. Its headline model, Olmo 3 Think 32B, is a "thinking" model: it writes out step-by-step reasoning before its final answer, as described in Extended Thinking. Ai2 said it was the first fully open thinking model of that size, and that its reasoning could be traced back to the training data and decisions that produced it. The release listed every stage: the 5.9-trillion-token pretraining mix, the smaller mixes used later in training, the post-training data (called Dolci), and checkpoints from each major milestone. Ai2 said the models were trained on up to 1,024 NVIDIA H100 chips.
On Ai2's own model card, Olmo 3 Think 32B lands close to Alibaba's Qwen 3 32B, ahead on some tests and behind on others: for example 85.4 against 88.8 on MMLU, a broad general-knowledge test, and 72.5 against 70.9 on the 2025 AIME maths competition questions. Ai2 also said Olmo 3 was trained on roughly six times fewer tokens than Qwen 3. These are the maker's own numbers on tests it chose, so read them as a claim, not a ranking.
An everyday case: a teacher shows a class what a model has read
Picture Amara, a secondary school computing teacher. Her students assume chatbots "look things up" or "just know". She opens the Ai2 Playground on the classroom screen, asks the model to explain photosynthesis, then presses the OlmoTrace button. Highlighted phrases in the answer turn out to match passages on real web pages in the training data. The class sees that the model's wording has roots in text people wrote, and that some of the matching documents are good sources and some are not. Then she asks a question with a wrong assumption built in, and the class watches the model go along with it. Nobody leaves thinking the model is an oracle, and nobody's personal data went into the chat, because she followed the playground's rule not to enter any.
Try it yourself
Buttons and model names change, so follow the steps in spirit if the labels differ.
- Open the Ai2 Playground at playground.allenai.org. At the time of writing it is free, and chatting did not appear to need an account (there is an optional sign-in). A box asks you to accept Ai2's terms. Read what it says: Ai2 may use your conversations to improve its models, and there are optional boxes about publishing your chats in public research datasets. Leave those optional boxes unticked unless you are happy to contribute. Never type personal or confidential information.
- Check which model is selected. The menu at the top shows the model name, such as an Olmo 3.1 model. Choose an Olmo model rather than one of Ai2's other models.
- Ask something where wording is likely to be borrowed. Try this prompt: “
Explain in three short paragraphs how vaccines train the immune system.“ - Press "Show OlmoTrace" under the answer. Look at the highlighted spans and open two or three of the matching documents. Ask yourself: is the source reliable? Did the model copy a phrase, or just a common expression?
- Test a known weakness. Ask a question whose answer changed recently, for example: “
What is your knowledge cutoff date, and who won the most recent football World Cup?“ Compare the answer with a trusted news site. Models only know what was in their training data, and may state an old fact confidently. - Try the same prompt in another model you use, such as ChatGPT or Claude. Notice which is more capable, and which lets you see where its words came from. That trade-off is the heart of this page.
- Optional, for the curious: run it on your own computer. Install the free Ollama app from ollama.com, open a terminal, and type
ollama run olmo-3:7b-instruct. At the time of writing that download is about 4.5 GB and needs a reasonably modern computer with plenty of memory. See Parameters for what size of model your machine can handle. - Optional: look at the data. Search Hugging Face for "allenai dolma" and open a dataset page. You can preview real training documents, which few other model makers let you do.
Common mistakes and misunderstandings
Thinking "open" always means the same thing. A downloadable model is not automatically open source. Ask three questions: Can I get the weights? Under what licence? Can I see the data and code used to make it? OLMo answers yes to all three; most open-weights families answer yes to the first only.
Assuming fully open means most capable. It doesn't. OLMo's largest models have 32 billion parameters, far smaller than the biggest commercial models. Ai2 compares its models mainly with open models of similar size and with older, smaller commercial models, such as GPT-3.5 Turbo and GPT-4o mini for OLMo 2 32B, not with the newest frontier systems. For hard real work, a leading closed model will usually do better. OLMo's strength is being inspectable.
Believing open data means clean data. Most of Dolma comes from web pages gathered by Common Crawl, a public web archive, so it contains the web's errors, biases, and writing whose authors never agreed to its use. Publishing it makes those problems visible rather than removing them. One Dolma 3 dataset card notes that some science documents were later redacted and marked "[REMOVED]" after training.
Treating the licence as permission for anything. The models use Apache 2.0, which allows commercial use, and the data uses ODC-BY, which asks for credit. But Ai2's model cards say the models are intended for research and education under its Responsible Use Guidelines, and third-party content in the data may carry its own rights. If you plan a business use, read the current terms yourself.
Expecting built-in safety like a commercial chatbot. Ai2's model cards warn that the models "can easily be prompted by users to generate harmful and sensitive content". Downloaded versions have few guardrails. Use them with care, especially with young people.
Reading OlmoTrace as proof of cause. A highlighted match shows the same words exist in the data. It doesn't prove the model learned the answer from that exact page.
Expecting strong non-English answers. The playground itself notes that Olmo is primarily designed for English.
Related elements
- Parameters: what model weights are, and how to run an open model on your own computer.
- Pretraining: the huge first stage of training that OLMo documents in unusual detail.
- Datasets: why the training data shapes a model, and why publishing it matters.
- Interpretability: the research effort to understand what happens inside models, helped by open data and checkpoints.
- Benchmarks: the tests behind every "beats model X" claim, and the contamination problem open data helps check.
- Llama: Meta's open-weights family, the main point of comparison for what "open" means.
- Qwen: Alibaba's models, many under Apache 2.0, and the rival Olmo 3 measures itself against.
- RLHF: the finishing training stages that Ai2 published openly as the Tülu recipe.
- Extended Thinking: the step-by-step reasoning that Olmo 3 Think writes out.
- Hallucination: confident wrong answers, which tools like OlmoTrace help investigate.