Olmo 3
by Allen Institute for AI (Team Olmo)
The technical report for the strongest fully-open thinking model released to date — every stage, checkpoint, and data point of the model flow documented.
Overview
Submitted to arXiv on December 15, 2025 and revised April 14, 2026, 'Olmo 3' is authored by Team Olmo — 65+ contributors including Allyson Ettinger, Luke Zettlemoyer, Noah A. Smith and Hannaneh Hajishirzi — and is filed under Computation and Language (cs.CL) and Machine Learning (cs.LG). The report introduces Olmo 3, a family of state-of-the-art, fully-open language models at the 7B and 32B parameter scales, with model construction targeting long-context reasoning, function calling, coding, instruction following, general chat and knowledge recall. Its defining contribution is the release of the entire 'model flow' — the full lifecycle of the family, including every stage, checkpoint, data point and dependency used to build it — so the models can be studied and retrained rather than merely used. The flagship Olmo 3 Think 32B is described by the authors as the strongest fully-open thinking model released to date. For practitioners, the report is the reference for how base, instruct, thinking and RL-Zero variants relate to one another and what training decisions produced each.
At a Glance
- Topic
- Models
- Level
- Advanced
- Format
- Paper
- Cost
- Free
- Duration
- Self-paced technical report
- Provider
- Allen Institute for AI (Team Olmo)
- Hands-on
- No
- Certificate
- None
What You’ll Learn
- ✓How a fully-open 7B/32B model family is constructed stage by stage, from base through instruct, thinking and RL-Zero variants
- ✓What the 'model flow' concept means in practice — releasing every checkpoint, data point and dependency, not just weights
- ✓How training is targeted at long-context reasoning, function calling, coding and instruction following
- ✓How an open thinking (explicit reasoning-chain) model is trained and evaluated against closed alternatives
- ✓Which artifacts you need to reproduce or extend a frontier-class open model yourself
Highlights
- •Flagship Olmo 3 Think 32B billed as the strongest fully-open thinking model released to date
- •Entire training pipeline, data, and intermediate checkpoints published alongside the paper
- •65+ authors from Ai2 and collaborators; revised as recently as April 2026
- •Free on arXiv, with the model family openly downloadable
Who It’s For
Best For
- ✓AI/ML engineers evaluating open models for self-hosting or fine-tuning
- ✓Researchers who need a reproducible reference training pipeline
- ✓Post-training and RL engineers studying reasoning-model recipes
- ✓Teams building on fully-open weights for licensing or auditability reasons
Prerequisites
- •Familiarity with transformer language models and LLM training terminology
- •Understanding of post-training methods (SFT, RL) helps
- •Comfort reading a long technical/academic report
FAQ
What is Olmo 3?
The official technical report for Olmo 3, Ai2's family of fully-open 7B and 32B language models, including the flagship Olmo 3 Think 32B reasoning model. Essential reading for AI engineers and researchers who want a complete, reproducible account of how a modern reasoning model is built rather than a weights-only release.
Is Olmo 3 free?
Olmo 3 is free to access.
What level is Olmo 3 for?
Olmo 3 is aimed at a advanced audience. Recommended background: Familiarity with transformer language models and LLM training terminology, Understanding of post-training methods (SFT, RL) helps, Comfort reading a long technical/academic report.
How long does Olmo 3 take?
Expect roughly Self-paced technical report. Most learners work through it at their own pace.
What will I learn from Olmo 3?
You'll learn: How a fully-open 7B/32B model family is constructed stage by stage, from base through instruct, thinking and RL-Zero variants; What the 'model flow' concept means in practice — releasing every checkpoint, data point and dependency, not just weights; How training is targeted at long-context reasoning, function calling, coding and instruction following; How an open thinking (explicit reasoning-chain) model is trained and evaluated against closed alternatives; Which artifacts you need to reproduce or extend a frontier-class open model yourself.