Ai2 releases OLMo 2 32B

Background

Most of the strongest language models reach users as a service. You can send them text and read what comes back, but the data they were trained on, the code that trained them and the intermediate states of the training run stay inside the organisation that built them. A "fully open" release is a different arrangement: together with the finished model, whoever released it publishes what another team would need to rebuild it. The difference matters for research, because a result nobody else can reproduce cannot really be checked.

What was released

On 13 March 2025, Ai2 released OLMo 2 32B ^[1]. The number in the name is the parameter count — roughly 32 billion weights, the numbers inside the network that training adjusts. That makes it a large model by the standards of openly released ones, though much smaller than the biggest systems in use, whose sizes are generally not published.

How it was trained

Ai2 says the model was trained on up to 6 trillion tokens ^[1]. A token is the unit these models read and produce, usually a fragment of a word; 6 trillion of them is an enormous amount of text, and the token count is one of the main dials that sets how long and how expensive pre-training becomes. The run itself took place on 1,280 H100 GPUs on Google Cloud's Augusta cluster ^[1]. The H100 is Nvidia's data-centre GPU, the chip most large training runs of this period are built around; at this scale the practical limit is usually how many of those chips can be kept busy on a single job at once, so the size of the cluster is part of the method, not just a logistics detail.

After pre-training, the model was post-trained using Tulu 3.1 ^[2]. Post-training is the second stage of building a chat model: a base model that can only continue text is taught, with curated examples and feedback, to follow instructions, answer questions and hold a conversation. Tulu is Ai2's own line of post-training data and recipes, so OLMo 2 32B was finished with a method the same group had published.

Results

Ai2 called OLMo 2 32B the first fully open model to beat GPT-3.5-Turbo and GPT-4o mini on a suite of academic benchmarks ^[1]. Two parts of that sentence are doing different work. The comparison sets a bar: GPT-3.5-Turbo and GPT-4o mini are smaller, widely used models, and they are the kind of practical baseline a new model has to clear to be useful. The qualifier sets another bar: the claim is not that OLMo 2 32B is the strongest model available, but that among models released with everything needed to reproduce them, it was the first to come out ahead on those tests. Academic benchmarks are standardised question sets — subject knowledge, reading comprehension, simple reasoning — that make models comparable, but they measure narrow, easily scored skills.

Ai2 also said the model cost about one third of what training Qwen 2.5 32B cost, while reaching similar performance ^[2]. Both models carry the same nominal size, so the comparison is between two runs of comparable scale, and the difference lies in what it took to get there. Training-cost figures are hard to compare across groups, because they depend on what is counted — chips, electricity, staff time, failed attempts — but a claim of this kind is exactly the sort of thing that a reproducible release lets outsiders examine.

Why it matters

Ai2 released everything needed to reproduce OLMo 2 32B ^[1]. This is where the "fully open" label gets its weight. For a research group it means being able to rerun the recipe, swap one ingredient and see what changes; for a student it means a model large enough to be interesting whose insides can be studied rather than only guessed at from the outside. It also turns comparisons such as the one-third cost claim into something that can be checked rather than merely stated. A model of this size, trained at this scale and published with its full recipe, is a reference point: it shows what a given budget of chips and tokens buys, and it gives everyone else a baseline against which the next open release can be measured.