Olmo 3

A model is the end of a series of choices

Training happens in stages. Pretraining builds a base model from a broad body of text. Midtraining continues on a more targeted mix, such as more mathematics and code, or on longer documents to extend the context. Post-training shapes behaviors such as following instructions and reasoning. Each stage starts from the weights the previous stage produced.

The November release includes Base and reasoning-focused Think models at 7B and 32B, plus two 7B branches: Instruct, for direct chat, and RL-Zero, described below. A 32B Instruct model arrived with Olmo 3.1 in December. [1] The architecture is a standard dense transformer. What is new is how much of the development path can be inspected.

Imagine two students sitting the same final exam. Their scores alone do not tell you whether a new textbook helped. Knowing where each started, which exercises they did and how they scored midway makes the comparison far more informative. Intermediate checkpoints play that role in training experiments. Repeated runs are still needed to tell a real effect from random variation.

Changing the training mix

The technical report documents the midtraining mixture, including mathematics, code and other data. It also studies how choices made early in training affect what reinforcement learning can achieve later. [2] That is a question about cause and effect: did the base model’s training data make the behaviors that reinforcement learning rewards easier to find?

The figure uses a made-up budget of 100 billion tokens. At a 30% targeted share, 30 billion tokens come from targeted data and 70 billion from broad replay data, general text mixed back in so the model keeps its general abilities. Raising the share to 60% keeps the total fixed while changing what the model sees. That controls one variable, the amount of data. Others stay uncontrolled: quality, duplication, difficulty and overlap with test sets also matter.

Predict what happens to the replay budget before moving the slider. The arithmetic gives the split exactly. It cannot say whether the model will get better at mathematics, worse at conversation, or both. Only measurements on held-out tests can answer that.

Several models from one base

One base model can be trained further along different paths. A reasoning path emphasizes long solutions and problems with checkable answers. A chat path emphasizes direct answers, following instructions and using tools. The RL-Zero branch applies reinforcement learning directly to the base model, skipping the supervised and preference-tuning stages that the Think branch goes through first. Because all the branches start from the same base, comparing them avoids many of the differences that make comparisons between unrelated models hard to interpret. [2]

The 32B Think model card lists its base model, training stages, data and released artifacts. Its training includes supervised fine-tuning, preference tuning and reinforcement learning. [3] “Think” names a model trained this way; it is not a different kind of network.

Suppose a reasoning model improves its math score but also answers ordinary questions more slowly. Measure those separately: correct answers per problem, tokens generated per attempt and total time. A single benchmark number cannot tell you whether the trade-off is worth it for a particular use.

Publishing is the first step toward reproducing

Public weights let a researcher run a model. Public data and code help explain how it was made and can make it possible to make it again. Ai2 documents how to run the models and publishes its training code through linked projects. [4] Reproducing a run still requires the computing power, the exact software versions, the same data processing, and tolerance for small numerical differences.

A useful study can be much smaller than retraining everything. Start from a released intermediate checkpoint, change one data source, keep a test set untouched, and report the runs that failed as well as those that worked. A good release reduces how much hidden information such a study needs.

Ai2’s December 2025 Olmo 3.1 update continued these training paths. Its checkpoints are different models from the November Olmo 3 releases and should be named as such in comparisons. [1] Exact model names and checkpoint identifiers are part of the evidence.

Simplifications and limits

• The mixture figure is a budgeting exercise, not a measured relationship between data and quality.
• Open data and code do not make a full reproduction cheap or numerically identical.
• This page covers the November Olmo 3 release, not the December Olmo 3.1 update.

Follow the idea

Tülu 3DeepSeek-R1

Sources and further reading

Primary papers, model cards, and release posts. Facts describe the versions named here; the interactive calculations are teaching models. Research cutoff: October 3, 2026.

1. Olmo 3 release and model flowNovember 20, 2025 release; later December 3.1 updates are explicitly separated.
2. Olmo 3 technical reportSections on pretraining, midtraining, model branches, and RL-Zero experiments.
3. Olmo-3-32B-Think model cardBase provenance and SFT/DPO/RL training artifacts.
4. Ai2 model documentationInference access and links to released model and training infrastructure.