OLMo by AI2 (Allen Institute for AI): status and timeline

Milestones

No announced next step

OLMo 1 released under Apache 2.01 Feb 2024Complete.
OLMo 2 released with 7B and 13B models26 Nov 2024Complete.
OLMo 3 family announced: 7B, 32B base and thinking variants20 Nov 2025Complete.
OLMo 3 32B Think released with 65K context window21 Nov 2025Complete.

Most important updates

AI2 announces OLMo 3: first fully open thinking model family from 7B to 32B20 Nov 2025

Current obstacles

Competing with proprietary labsOpen-source development faces stiff competition from well-funded frontier labs that don't publish their work.
Scaling reasoning efficientlyMaking thinking models competitive with frontier models while remaining practical to run.

Physics limits

Attention memory scales with sequence lengthIn standard Transformers, storing attention weights grows quadratically with context length, limiting the window size and requiring architectural changes for very long contexts.
Reasoning compute cannot be compressedStep-by-step reasoning requires more forward passes through the model, so thinking models inherently cost more to run per query than baseline inference.

How it works

3 parts

Training data

Dolma corpus

OLMo trains on Dolma, a publicly documented corpus of internet text. Dolma 3 adds curated scientific and technical content to improve reasoning.

Thinking mode

Step-by-step reasoning

OLMo 3 32B Think generates explicit reasoning chains before answering, showing the model's step-by-step logic.

Transparency

Full source code and weights

Training code, model weights, data processing pipelines and checkpoints are all published under Apache 2.0.

Spec sheet

Spec	OLMo
Context window	65,536 tokens^R (reported)
OLMo 3 32B Think parameters	32 billion^R (reported)

^R reported by the company

Papers & demos

1 item

1 Feb 2024code
OLMo 1 released (opens allenai.org)First major open-weight language model from Ai2, setting the standard for reproducible, fully transparent model development.

Update log

1 update

Thu 20 Nov 2025

Major: AI2 announces OLMo 3: first fully open thinking model family from 7B to 32BPressOther

About AI2 (Allen Institute for AI)

The team behind OLMo

AI2 (Allen Institute for AI)

OLMo, open-weight reasoning models, AI research nonprofit

AI2 is a nonprofit research institute founded by Paul Allen in 2014. It develops fully open-weight language models like OLMo 3, providing complete transparency in weights, code and training data under Apache 2.0.

Paul Allen's nonprofit pursuit of AI research as a public good, operating since 2014 with teams in Seattle and Tel Aviv.
OLMo 3, released November 2025, is the first fully open 32B thinking model with 65,536 token context window, trained on Dolma 3 corpus.
Ai2 provides complete source transparency: weights, training code, datasets and checkpoints under Apache 2.0 license.

allenai.org (opens allenai.org)Company profile →

Founded: 201412 yrs
Headquarters: United States
Status: Private
Valuation: private
Works in: AIOpen-weight models
Coverage: 1 program · 1 updatelatest 20 Nov 2025checked 10 Oct
People: Peter Clark (opens allenai.org)Interim CEO (from March 2026)Oren EtzioniFounding CEO (2014–2022)

Spec

OLMo

Context window

65,536 tokens^R (reported)

OLMo 3 32B Think parameters

32 billion^R (reported)