AI Potluck
Contribute
Back to Gap Map Model components / Fine-tuned / chat models

OLMo 3 Instruct

Allen Institute for AI
open source / Overall score: 3.7

Instruction-tuned chat releases of Ai2's OLMo 3 family at 7B and 32B, post-trained on the fully open OLMo 3 base with the open Tulu and OLMo recipe. Every stage, checkpoint, dataset and dependency needed to recreate or modify them is published. It supersedes OLMo 2 Instruct.

Adoption counts the two declared instruct checkpoints, 7B and 3.1 32B, and not their SFT, DPO or quantized variants, so it is a floor.

Openness

5 high confidence
5.0
weights
open(Apache-2.0)
data
open(Dolma 3 pretraining + open post-training mixtures)
code
open(OLMo + open-instruct)
checkpoints
open
license
Apache-2.0(OSI)

A fully open instruct line: Apache-2.0 weights, open pretraining and post-training data, open training code and open checkpoints. The line now extends to an Olmo 3.1 update shipping 32B Think and Instruct checkpoints on the same open recipe, which does not change the class.

  • https://allenai.org/blog/olmo3 recorded 2026-09-29

    The post defines the model flow as "every stage, checkpoint, dataset, and dependency required to create and modify it", links Open Instruct for code, and says "Together, Dolma 3 and Dolci give Olmo 3 a fully open data curriculum".

  • https://huggingface.co/allenai/Olmo-3-7B-Instruct recorded 2026-09-29

    The model card is licensed under "Apache 2.0. It is intended for research and educational use" and its embedded repo state carries "gated":false, so the weights download without a barrier.

Adoption

3 medium confidence
3.0

Adoption is measured as Hugging Face downloads summed across the two declared releases, Olmo-3-7B-Instruct and Olmo-3.1-32B-Instruct. The 32B checkpoint declared before it, Olmo-3-1025-32B-Instruct, is no longer on the Hub, and 3.1 is the 32B Instruct Ai2 publishes now.

Capability

4 high confidence
4.0

Ai2's evaluations show Olmo 3 Instruct tying or surpassing Qwen 2.5, Gemma 3 and Llama 3.1 at similar scale and competing with the Qwen 3 family, and the 3.1 32B Instruct is described by Ai2 as the strongest fully open 32B-scale instruct model. It is a fully open line that stays below the current frontier of proprietary and largest open-weight models.

  • https://allenai.org/blog/olmo3 recorded 2026-09-29

    The post reports "Olmo 3-Instruct ties or surpasses Qwen 2.5, Gemma 3, and Llama 3.1 in our evaluations, and competes with the Qwen 3 family at similar scale".

Verified 2026-09-29, except some scores, still awaiting re-confirmation