Tier 5 Fully open / reproducible
What the publisher claims: fully open, open-source, open weights, open data
The one-liner: Public-institution fully open model with strong provenance goals.
The six questions
Weights for 1.0, the 1.1 mini distills, and the multimodal 1.5 are downloadable, together with base, SFT, RLVR, and DPO checkpoints and the curated data collections.
Weights, code, and data are under Apache 2.0, but downloading requires accepting an Apertus Acceptable Use Policy that carries obligations Apache 2.0 does not, including an indemnity to ETH Zurich and EPFL and a trademark restriction; the grant is permissive with caveats.
Independently deployable at 8B and 70B and at the smaller distilled sizes, with standard open inference stacks and third-party hosting by Swisscom, Infomaniak, and others.
A data transparency repository publishes every training dataset with reproducible versions and licenses; the 1.5 technical report documents recipes, ablations, and data mixtures; all 95 evaluations are published in the Every Eval Ever schema so anyone can rerun them.
Independent groups can rebuild 1.5 from published data, code, and recipes and reproduce the published evaluations, which the EvalEval Coalition treats as the reference case for reproducible evaluation reporting.
Published governance structure: steering committee, consensus-driven decisions, regular community meetings, and the swiss-ai org shows 110 merged outside PRs on evals-post-train alone.
License
Apache 2.0 for source code, weights, and data, accompanied by a mandatory Apertus Acceptable Use Policy and an indemnity to ETH Zurich and EPFL — license source ↗
- Hugging Face downloads require accepting an Acceptable Use Policy that adds obligations absent from Apache 2.0
- Users agree to indemnify, defend, and hold harmless ETH Zurich and EPFL against third-party claims
- APERTUS is a registered trademark and may not be used to imply endorsement
What is actually released
- weights (8B, 70B, 0.5B/1.5B/4B mini distills, multimodal 1.5)
- intermediate checkpoints (base, SFT, RLVR, DPO)
- full training data with reproducible dataset versions and per-dataset licenses
- training code and recipes (technical report)
- alignment principles (Apertus Charter)
- evaluations published in the Every Eval Ever schema with rerunnable harness
Why it matters
Apertus is the sovereign and public-institution comparison to corporate open-weight releases: the only major non-commercial model family publishing its data, its recipes, and its evaluation protocol.
Sources
- https://www.apertus-ai.org
- https://www.apertus-ai.org/pages/about
- https://apertus-ai.org/docs/tech/licensing
- https://www.apertus-ai.org/articles/2026-07-apertus-1-5
- https://arxiv.org/html/2610.15542v1
- https://evalevalai.com/infrastructure/2026/10/15/apertus-open-reproducible-evals
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.