Open Models
Open weights you can run on your own infrastructure — Swedish-first chat, embeddings, and reranking, published on Hugging Face.
Model lineup
Chat, continued fine-tuning, and a Swedish retrieval stack — all designed for local and on-prem deployment.
Swedish-enhanced European assistant (Gemma 2 9B). Run locally with Ollama.
Light QLoRA fine-tune for Swedish assistant behavior and European context. Ships as GGUF Q4_K_M (~5.5 GB) with an Ollama Modelfile — three commands to chat.
LoRA adapter for continued fine-tuning (advanced).
The PEFT adapter behind Europa v1.0 — for further training or serving via Transformers + PEFT instead of the merged GGUF.
Swedish-first EuroBERT embedding model for semantic search and RAG.
Maps queries and documents into a shared 768-d space for fast candidate retrieval. Pair with the reranker for higher precision.
Swedish-first EuroBERT cross-encoder for reranking retrieved documents.
Scores query–document pairs after first-stage retrieval. Use on top candidates from diabase-embedding-1 before grounding Europa.
Quick start
Europa ships as a quantized GGUF with a ready-made Modelfile. Accept the Gemma license on Hugging Face if required by your org, then create and run the model locally.
Recommended: ~6 GB VRAM for GPU, or 16 GB RAM for comfortable CPU inference.
Full model card
# 1. Download model files
huggingface-cli download diabase/europa-9b-GGUF \
gemma-2-9b-it.Q4_K_M.gguf Modelfile
# 2. Create the local model
cd <download-folder>
ollama create diabase-europa-9b -f Modelfile
# 3. Run
ollama run diabase-europa-9bRetrieval stack
A simple pipeline for Swedish RAG without sending documents to a third-party API.
diabase-embedding-1 retrieves a broad candidate set from your corpus.
diabase-reranker-1 scores those candidates for higher precision.
Diabase Europa answers from the grounded context — locally.
Why open models
Closed models mean closed infrastructure. You can't audit them, you can't run them on your own hardware, and you can't be sure where your prompts end up.
I publish open weights, transparent evaluation, and clear limitations. Early-stage models are labelled as research previews — no overclaiming.
Licenses & attribution
Each model ships with clear license terms. Credit where it's due.
Europa 9B and the LoRA adapter are derived from google/gemma-2-9b-it and are subject to the Gemma Terms of Use, including attribution requirements.
Diabase fine-tuning artifacts and documentation are Apache 2.0 where applicable. Derivative works remain subject to the Gemma license.
diabase-embedding-1 and diabase-reranker-1 are released under Apache License 2.0, consistent with the EuroBERT foundation.
Both are research previews. Evaluate them against your own Swedish corpora before production use.
These models build on open foundations and tooling:
All models are on Hugging Face under the Diabase org. Download them, run them locally, and tell me what you build.