Olmo 3

What is Olmo 3?

Olmo 3 is the third generation of fully open language models from Ai2, the Allen Institute for AI. It comes as dense 7B and 32B models in Base, Think (reasoning) and Instruct (chat and tool use) versions, plus RL Zero research checkpoints. Where Qwen or Gemma release only weights, Ai2 also publishes the training data, code and checkpoints from every stage, so you can study, fork or retrain it.

The Olmo 3.1 update added stronger 32B Think and Instruct models, and Ai2 still lists Olmo 3 as its latest Olmo family. It works in English, reads up to 65K tokens, and its knowledge stops at December 2024.

What changed in Olmo 3.1?

Two stronger 32B models and refreshed research checkpoints, under the same Apache 2.0 licence.

32B Think

Ai2 ran reinforcement learning for 21 more days on 224 GPUs, lifting AIME 2025 from 72.5 to 78.1 and IFBench from 47.6 to 68.1.

32B Instruct

A new 32B chat and tool-use model; Olmo 3 had Instruct only at 7B, and Ai2 calls this its most capable fully open chat model.

RL Zero

New 7B Code and Math checkpoints from longer, more stable training runs, for reinforcement learning research.

Playground

Olmo 3.1 32B Instruct is now the default model in the free Ai2 Playground.

Who is Olmo 3 best for?

Olmo 3 is best for AI research, custom fine-tuning, audits that need known training data, and local English chat with tool use.

• AI research

  Checkpoints from pretraining, mid-training, long-context training and each post-training stage let you study how skills appear or rerun a single stage.

• Custom fine-tuning

  Start from the Base model or any earlier checkpoint under Apache 2.0, and reuse Ai2's own SFT, DPO and reinforcement learning recipes.

• Audits that need known data

  The Dolma 3 pretraining data and the Dolci post-training sets are downloadable, and OlmoTrace shows where answer text matches the training data.

• Local English chat and tool use

  The 7B Instruct model handles multi-turn chat and function calling, and Ai2 built the family to run on hardware down to laptops.

What's included with Olmo 3?

Think and Instruct versions, checkpoints from every training stage, open datasets, OlmoTrace, function calling and open training tools.

Think and Instruct versions

Think models write out step-by-step reasoning; Instruct models give shorter answers and call tools.

The full model flow

Ai2 publishes checkpoints from every training stage, not only the final weights.

Open training data

Dolma 3 (a 9.3T-token pool with a 5.9T-token training mix), the Dolmino and Longmino mixes and the Dolci post-training sets are free to download.

OlmoTrace

In the Ai2 Playground, it highlights spans of an answer that exactly match documents in the training data.

Function calling

Instruct models return tool calls through vLLM's olmo3 parser on an OpenAI-compatible endpoint.

Open training tools

OLMo-core, Open Instruct, OLMES and decon cover training, post-training, evaluation and removing test sets from data.

How much does Olmo 3 cost?

Olmo 3 is free: the weights and training data download at no cost under Apache 2.0, and the Ai2 Playground is a free web chat; Ai2 sells no paid plan.

Ai2 sells no paid plan or API of its own for Olmo. Hosted API access comes through Ai2's inference partners, such as OpenRouter, at their own prices. Full pricing on playground.allenai.org. We checked these prices on 28 Sep 2026.

How do readers rate Olmo 3?

No one has rated Olmo 3 yet. Be the first.

Olmo 3 pros and cons

It is the most open model family you can use, with strong maths and instruction following, but it trails Qwen at the same size, runs slower than mixture-of-experts models and reads only 65K tokens of English.

Is Olmo 3 open source?

Yes. Olmo 3's code is public under the Apache 2.0 licence, so you can run it on your own server.

What are the best Olmo 3 alternatives?

Olmo 3 is an open source alternative to Falcon H1R 7B. Qwen3.5 Small and Gemma 4 do the same job and are open source too.

Falcon H1R 7BFree

TII's downloadable hybrid reasoning model for maths, code and logic.

Qwen3.5 SmallOpen Source

Qwen3.5's 0.8B to 9B open multimodal models.

Gemma 4Open Source

Choose it when you need image or audio input, a 128K to 256K context, or sizes small enough for phones; Olmo 3 reads English text only.

Qwen3.6-27BOpen Source

Choose it when you want stronger coding and image or video input at a similar size and do not need open training data; it reads 262K tokens natively.

Nemotron 3.5 LightningOpen Source

Choose it when you want another model with open weights, data and recipes that runs faster; it is a 30B mixture of experts with 3B active and a 1M-token context.

Questions about Olmo 3