Apertus 1.5 vs OLMo 3.1: which fully open model should you use?
Apertus 1.5 reads text, images and audio with a 262K context; OLMo 3.1 is text-only with 64K and has a 32B model. Both are fully open under Apache 2.0.
Verdict. Both are fully open under Apache 2.0, with training code and checkpoints public, so choose by what the model must read and how recent it needs to be. Apertus 1.5 (8B and 70B, July 2026) takes images and audio as well as text, has a 262K context and optional thinking, and was trained only on openly available data. Its repositories are gated, audio understanding is experimental, and it needs Swiss AI's forks of Transformers or vLLM. OLMo 3.1 (7B and 32B, December 2025) is text-only with a 65,536-token context, and Ai2 publishes the pre-training data (Dolma 3) and the post-training datasets (Dolci) along with the code and checkpoints. Ai2 calls the 32B Instruct model the strongest fully open 32B instruct model in its evaluations. OLMo is older and trails the 2026 open-weight models on capability, while Apertus is newer and larger, with a 70B model.
Specifications
| Specifications | Apertus 1.5 | Olmo 3.1 |
|---|---|---|
| Developer | Swiss AI Initiative | Allen Institute for AI (Ai2) |
| Released | July 24, 2026 | December 12, 2025 |
| Total parameters | 70B (Apertus 1.5 70B) | 32.2B (32B Think) |
| Active parameters | All | All |
| Architecture | Dense | Dense |
| Context window | 256K tokens | 64K tokens |
| Input | Text, Image, Audio | Text |
| License | Apache 2.0 (Permissive) | Apache 2.0 (Permissive) |
| Memory for the smallest variant (4-bit) | 5.52 GB (Apertus 1.5 8B) | 5.04 GB (7B Think) |
Performance claims come from the developer and have not been independently verified.
Pick Apertus 1.5 if
- You need a model that reads images or audio as well as text.
- You need a model larger than OLMo's 32B, since Apertus has a 70B version.
- You want a 262K context and thinking as a chat-template flag.
- You must document where the training data came from: Apertus uses only openly available data and honors crawler opt-outs retroactively.
Pick Olmo 3.1 if
- You want to study or reproduce training with the actual data, since Dolma 3 and Dolci are published.
- You want a 7B Think or Instruct model of about 15 GB to run locally.
- You want to run reinforcement-learning experiments from the released RL-Zero and intermediate checkpoints.
- You work with text only and want public data and checkpoints for every stage: base, supervised fine-tuning, preference tuning and reinforcement learning.