Apertus – Open Foundation Model for Sovereign AI
An open Swiss foundation model project, Apertus, is prompting debate over what “sovereign AI” should mean in practice: fully open weights, data, and training pipelines versus reliance on closed, US‑controlled frontier systems. Commenters welcome Apertus and similar efforts from Europe and China for enabling independent, inspectable models, but many question Apertus’s current quality, data ethics, and pace relative to stronger open alternatives like Nemotron, OLMo, K2, and Chinese LLMs. A broader thread runs through the exchange about data privacy, geopolitical risks of US tech dominance, and whether the real near‑term battle is not just open vs closed models, but cloud AI services vs usable local models on consumer hardware.
Overall focus
- Discussion centers on Apertus as a “fully open” model (weights, data, training recipes) aimed at European/Swiss “sovereign AI” and whether that matters given its current capability.
Openness, pipelines, and “SOTA”
- Many value Apertus for being genuinely open: open weights, open data, full training pipeline.
- Some argue true “state of the art” should mean models that can be inspected and replicated, not closed “cutting-edge” systems from frontier labs.
- Others maintain that frontier lab models remain the real performance SOTA, regardless of openness.
Model quality and practical use
- Earlier Apertus versions were described as “pretty bad”; some testing suggests the new ones are still not competitive with top models.
- Users report it’s workable as a backbone for RAG and some agents (e.g., legal consulting, translation), but not yet “agentic” or frontier-level.
- Weaknesses include hallucinations in multilingual tasks and basic language questions (e.g., conjugations, word spellings).
Training data, copyright, and ethics
- Apertus uses FineWeb/Common Crawl; some criticize this as unlicensed scraping that contradicts “copyright-compliant” marketing.
- Others argue scraping public web data for training is legal and that expanding copyright here would be harmful.
- There’s demand for a “vegan” model trained only on licensed or public-domain data for ethical reasons.
Sovereign AI, geopolitics, and data locality
- Strong theme: countries (especially in Europe) need their own AI capabilities to avoid dependence on US or Chinese tech, given concerns about US rule of law, surveillance, export controls, and political instability.
- Some see Apertus and similar projects as capability-building more than immediate model competitiveness.
- Debate over which jurisdictions are safest for data (US vs EU vs Switzerland vs Nordics) and whether any country is truly “safe.”
Comparison with other open models
- Other fully open or near-open pipelines mentioned: OLMo 3.1, K2 Think V2, Nvidia Nemotron, plus strong Chinese models (GLM, DeepSeek, Qwen).
- Consensus that Nemotron and several Chinese models currently outperform Apertus; some users prefer them in production.
Local vs service models and UX
- Several argue the real near-term battleground is local vs hosted LLMs, not just open vs closed.
- Local models are already “good enough” for many tasks, but tooling and UX are confusing and fragmented.
- Concern that poor local UX is pushing users toward centralized, closed services, reducing digital autonomy.
Compute, licensing, and compliance
- Claim that “the Swiss have no GPUs” is refuted by references to the Alps supercomputer with thousands of Grace-Hopper chips.
- License includes a novel mechanism: periodically downloading a hash-based filter to remove personal data from outputs based on deletion requests; unclear how sustainable this is.
- Some see Apertus mainly serving European compliance/sovereignty requirements rather than chasing peak benchmark scores.