LightOn defends its Europe-built models: DeepSeek-cited architecture, 70M ModernBERT downloads
IgorCarron · x · 2026-10-05
LightOn co-founder Igor Carron weighed in on the debate over Europe-built AI models.
- A quoted tweet notes a European model was trained from scratch, not fine-tuned from a Chinese checkpoint — but its MoE design, routing, and load balancing cite DeepSeek and Kimi papers, and most synthetic data came from Chinese models. Open source is a global commons, and that's fair game.
- Carron adds that even the French don't know LightOn: it built PAGnol, the first French LLM (1.5B then 10B), in 2021, all in Europe; its later models include ModernBERT with 70M cumulative HuggingFace downloads, and its OCR was ahead of US and Chinese models at release.
Related event: Overlooked French AI Veteran LightOn Hits 70M Open-Source Model Downloads(2 posts)→
More from Companies & People
- Codex Community Runs Weeklong Global Meetups Across 15+ Cities Post-DevDay — paw_lean · 2026-10-05
- Anthropic-Accenture 'embedded evaluation' sparks conflict-of-interest concerns — DavidLinthicum · 2026-10-05
- OpenClaw hype dying as corporations move in, dev laments loss of open-source community joy — alex_verem · 2026-10-05
- Qdrant and Neo4j to host Engineers Debate Night at SF Tech Week, debating 'open weights should be banned' — qdrant_engine · 2026-10-05
- UMass Amherst to hire tenure-track NLP faculty, meeting candidates at TADA/COLM 2026 — brendan642 · 2026-10-05
- India's Sarvam AI ships sovereign 30B/105B models plus a 22-language voice stack — CurieuxExplorer · 2026-10-05