Ethical Stack

Apertus-8B-2509

Original 8B 2025 open checkpoint, now superseded by Apertus 1.5 8B. It remains downloadable under Apache 2.0 and has a detailed public training card.

Score calculation

(7 + 7 + 9) ÷ 3 = 7.7

3 of six dimensions rated. Unrated dimensions and BDS status are excluded.

Full scoring method ↗
Setup details

Findings are specific to the 2509 checkpoint and its model card; current Apertus 1.5 improvements or safeguards are not attributed to this older release.

Swiss AI's EPFL lab page records Apertus 8B/70B model weights released September 2, 2025; official documentation now lists 1.5 as the current generation. Checked 2026-10-23.

Energy

↑

The 8B model card reports 15T pretraining tokens and full training details, but does not provide an 8B-specific GPU-hour, kWh, emissions, or water total. The technical report's approximately 6 million GPU-hours and estimated 5 GWh at 560 W per module apply to the 70B model's 15T-token main pretraining run, not this 8B checkpoint; the report says CSCS relies entirely on hydropower and uses closed-cycle Lake Lugano cooling with no water consumed. The 70B figures are not assigned to the 8B entry. No standardized inference-energy measurement is reported.

Reasoning & sources

The 8B card gives token scale and reproducible training details, while the report’s energy and hydropower figures apply to the separate 70B run. This shows training-process transparency, not the 8B checkpoint’s footprint; do not transfer the larger model’s numbers.

No exact 8B training energy/emissions or inference per-task/per-token measurement; the family report’s 70B estimate is not transferable to this release.

Training rights

↑

The model card says training data is open and compliant, respects data-owner opt-outs retrospectively, and links to data reconstruction scripts. These are developer-reported controls, not independent legal certification.

Reasoning & sources

The card says data are open and honors retrospective opt-outs, with reconstruction scripts available. These practices support a positive process judgment, but developer claims and scripts do not independently verify every rightsholder’s permission.

Published corpus and rights-handling disclosures are unusually detailed, but remain developer-reported and lack independent rights verification; score reflects transparency with that material limitation.

No independent audit of every source or rights claim.

Privacy

↑

Self-hosting lets the operator control prompts. The card states no PII output filter was provided at publication and advises users to apply the project's periodically updated filter; it also provides privacy/copyright request contacts.

Reasoning & sources

Local weights let operators control prompt storage, but the card says no PII output filter was supplied and recommends periodic opt-out updates. Deployment can protect input privacy while still requiring output controls and refreshed data handling.

Local inference is possible, but the card says no PII output filter was provided at release.

No centrally hosted service retention policy applies to all self-hosted uses; operator controls matter.

Openness

↑

Weights are published under Apache 2.0, with open data, training code, recipes, intermediate checkpoints, and training details described in the model card and project materials.

Reasoning & sources

Apache 2.0 weights, training code, recipes, data, and intermediate checkpoints provide unusually broad inspection and reuse. Full reproduction still depends on the linked data and environment, and independent verification of dataset rights is not available.

Weights, data, training code, recipes, and intermediate checkpoints are publicly described.

Confirm linked artifacts remain available and match the exact checkpoint revision.

Worker treatment

↑

the model card identifies training infrastructure and research contributors but does not disclose annotation worker compensation, labor conditions, or vendor chain.

Reasoning & sources

The model card identifies infrastructure and research contributors but does not describe annotation workers’ pay or conditions. The sources leave worker conditions unknown; ask about any human-curated corpora or safety work.

No worker-condition disclosure found.

Government & military

↑

The Apertus 8B 2509 card identifies a downloadable model release; the project’s published documentation describes Apache 2.0 as the family’s release license. Independent downstream use is possible, but no checked source identifies military adoption.

Reasoning & sources

Apertus distributes its checkpoint files and documents the family under Apache 2.0, which allows downstream deployment outside the original research organizations. That weakens any inference from vendor contract lists; no cited deployment record establishes military use of this checkpoint.

Needs specific downstream customer/procurement evidence.

BDS status

↑

We did not find Swiss AI Initiative (EPFL, ETH Zurich, CSCS) among the named targets in the two official BNC campaign lists. This is a limited list check, not BDS approval.

Reasoning & sources

Swiss AI Initiative (EPFL, ETH Zurich, CSCS) is not a named target in the checked lists, but BDS uses selective campaigns, so absence cannot establish a clean record. Running downloaded weights locally is a different decision from paying the developer or a cloud host; check the host as well.

This review covers the named targets in the BNC consumer/pressure guide and No Tech campaign, not every BDS campaign, local group, investment or supplier. Lists and ownership can change. A company’s absence from these lists is not an endorsement or proof that it has no relevant relationships.

All sources

Side by side