What would it take for AI to be open?

Michael Falk

The University of Melbourne

Fully open AI

Apertus is a fully open model. We pair the release of the weights of the Apertus model suite with a full set of reproduction artifacts, including source code, final and intermediate model checkpoints, reproducibility scripts for training data, evaluation suites, and this technical report. (Hernández-Cano et al. 2025, 7)

“Reproduction artifacts”

  1. Source Code ()
  2. Weights (🤗)
  3. Data (🤗, )
  4. Evaluation Code ()
  5. Documentation (arXiv)

How open is AI?

Open models discussed in the Apertus report (Hernández-Cano et al. 2025, 40)

How open can it be?

For the non-controversial prompt-completion pairs (Section 4.1.4 above), we assign rewards with a pretrained reward model. Specifically, we use Skywork-Reward-V2-Llama-3.1-8B (Liu et al., 2025a), an 8B-parameter Llama 3.1 decoder finetuned on 26M preference pairs curated with a human–AI annotation pipeline. As of summer 2025, it ranks highly on reward model benchmarks (Liu et al., 2025a).

… a suite of eight reward models ranging from 0.6B to 8B parameters, trained on a carefully curated subset of 26 million preference pairs from SynPref-40M. (Liu et al. 2025)

References

Hernández-Cano, Alejandro, Alexander Hägele, Allen Hao Huang, Angelika Romanou, Antoni-Joan Solergibert, Barna Pasztor, Bettina Messmer, et al. 2025. “Apertus: Democratizing Open and Compliant LLMs for Global Language Environments.” arXiv. https://doi.org/10.48550/arXiv.2509.14233.
Liu, Chris Yuhao, Liang Zeng, Yuzhen Xiao, Jujie He, Jiacai Liu, Chaojie Wang, Rui Yan, et al. 2025. “Skywork-Reward-V2: Scaling Preference Data Curation via Human-AI Synergy.” arXiv. https://doi.org/10.48550/arXiv.2507.01352.