European AI · Open weights
Open models and tools for teams that need data sovereignty — Swedish-first NLP, local deployment, and infrastructure you can actually inspect. Built by Mattis Rosén.
Why Diabase
Built for people who can't shrug about where prompts go or which law sits on top of the stack.
Prefer EU-hosted paths and self-hostable weights. Know where workloads run — and keep the option to keep them on your own metal.
In Chisel, restrict data handling and model use to a country or region you choose. You control the boundary — no silent exceptions.
European AI shouldn't mean slow AI. Lean stacks, practical quantization, and infrastructure close to the people using it.
Projects
Products and research a recruiter can inspect — write-ups first, live demos second.
Full-stack AI workspace: conversations, model integration, workflows, and region lockdown. Architecture and decisions without needing an account.
Project write-upEmbedding + reranker research previews: adaptation, contrastive training, evaluation leakage audit, and honest MTEB comparison.
Case studyEuropa 9B (QLoRA → GGUF/Ollama), LoRA adapter, and the full catalog on Hugging Face.
Model lineupData sovereignty
A lot of AI infrastructure still routes through US-based providers, under laws that sit awkwardly next to European privacy rules.
Diabase is my attempt at a practical alternative: open weights, local-first options, and tooling that doesn't force a trade-off between capability and control.
Principles
Less theatre, more operable systems.
Weights, licenses, and limitations on the table. Research previews stay labelled as such.
Ollama, GGUF, and stacks you can put on your own GPU — not only an API behind a login.
Language and jurisdiction aren't afterthoughts. Tuned for Swedish use, designed for EU context.
Get updates
Join the waitlist for launches and technical notes. Low volume — nothing spammy.
