Olmo 3.1
Ai2's fully open 32B and 7B models: weights, training data, code and checkpoints all public, Apache 2.0.
In short. Olmo 3.1 is a family of AI models from a non-profit research institute. What sets it apart is that almost everything behind it is published: not just the model weights, but the training data, the code and the intermediate checkpoints. That makes it the best choice for people who want to study or reproduce how a model is made.
Summary
The Allen Institute for AI (Ai2) released the Olmo 3 family on 20 November 2025 and added two stronger 32B models, Olmo 3.1 Think 32B and Olmo 3.1 Instruct 32B, on 12 December 2025. The family also includes 7B Think and Instruct models. All use the Apache 2.0 license.
"Fully open" is the point. Ai2 releases the weights together with the pre-training data (Dolma 3), the post-training datasets (Dolci), the training code and intermediate checkpoints, so that others can inspect and reproduce each stage: base model, supervised fine-tuning, preference tuning and reinforcement learning with verifiable rewards.
The models are dense transformers with a 65,536-token context window. Olmo 3.1 Think 32B was produced by continuing the reinforcement-learning run of Olmo 3 Think 32B for another 21 days on 224 GPUs; Ai2 reports gains of more than 5 points on AIME and more than 20 points on IFBench compared with Olmo 3 Think 32B. Olmo 3.1 Instruct 32B applies the recipe of the 7B Instruct model at larger scale and is aimed at chat, tool use and multi-turn dialogue. Ai2 calls it the strongest fully open 32B instruct model in its evaluations. These are the developer's own results.
The 32B BF16 checkpoints are about 64 GB, and the 7B ones about 15 GB. The family is text-only and was published in late 2025, so it predates the 2026 models in this catalog and trails the top open-weight models on raw capability.
Good for
- Research and teaching on how language models are trained, since data, code and checkpoints are open
- Reinforcement-learning experiments, using the released RL-Zero and intermediate checkpoints
- Organizations that need provenance and transparency about training data
- Running a dependable 7B or 32B model locally under Apache 2.0
Watch out for
- It trails the best 2026 open-weight models on capability; choose it for openness, not for peak benchmark scores
- Text only, with a 65,536-token context
- Released in late 2025; newer models in this catalog lead on coding and agent tasks
- The Think models write long chains of thought, which costs tokens and time
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the 7B Think variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| 7B Think BF16 | 7.3B | 17 GB | 8.39 GB | 5.04 GB | Laptop or phone |
| 7B Instruct BF16 | 7.3B | 17 GB | 8.39 GB | 5.04 GB | Laptop or phone |
| 32B Think BF16 | 32.2B | 74 GB | 37 GB | 22 GB | Gaming GPU |
| 32B Instruct BF16 | 32.2B | 74 GB | 37 GB | 22 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
- Apertus 1.5 vs OLMo 3.1: which fully open model should you use?
- Best small open models you can run on a laptop
- Apertus 1.5: Switzerland's fully open model reads images and audio
Frequently asked questions
- What is Olmo 3.1?
- Olmo 3.1 is a family of AI models from a non-profit research institute. What sets it apart is that almost everything behind it is published: not just the model weights, but the training data, the code and the intermediate checkpoints. That makes it the best choice for people who want to study or reproduce how a model is made.
- Can I use Olmo 3.1 commercially?
- Yes. Olmo 3.1 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Olmo 3.1 open source?
- Olmo 3.1 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Olmo 3.1 on my own computer?
- Yes. The smallest variant, 7B Think, needs about 5.04 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Olmo 3.1 released?
- Olmo 3.1 was released by Allen Institute for AI (Ai2) on December 12, 2025.
- How long a context window does Olmo 3.1 have?
- Olmo 3.1 supports a context window of up to 64K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Ai2 blog: Olmo 3, charting a path through the model flow (20 Nov 2025, updated 12 Dec 2025 for Olmo 3.1) allenai.org
- Olmo-3.1-32B-Think model card (Hugging Face) huggingface.co
- Olmo 3.1 collection (Hugging Face) huggingface.co
Release date and facts last verified November 8, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| Apertus 1.5 Swiss AI Initiative | 8B – 70B | 256K | Apache 2.0 (Permissive) | 5.52 GB Laptop or phone | |
| K2 Horizon Institute of Foundation Models (MBZUAI) | 1.08B – 379B 23B active | 512K | Apache 2.0 (Permissive) | 0.745 GB Laptop or phone |