A small open model that speaks good Norwegian and calls tools reliably, running on one 16 GB GPU
Majere is an independent, open project to fine-tune a 7–8B open-weight language model for Norwegian Bokmål and Nynorsk and for reliable function/tool calling, quantised so it runs locally on a single consumer GPU with 16 GB of VRAM.
The problem
- Small open models are weak in Norwegian, especially Nynorsk. Strong Norwegian models exist (e.g. NorMistral from the University of Oslo), but they are not trained for tool calling.
- Models that are good at tool calling are trained mostly in English. Asked in Norwegian, they get worse at picking the right function, filling in arguments (dates, addresses, amounts) and producing valid JSON.
- Running locally matters. Public bodies, schools and small businesses in Norway often can't send data to foreign cloud APIs. A model that fits on one 16 GB card can run on-premise.
What we will deliver
- Open LoRA adapters and merged 4-bit (GGUF/AWQ) models that run on a 16 GB GPU.
- An open Norwegian function-calling benchmark in Bokmål and Nynorsk, since none exists today as far as we have found.
- A fully reproducible training and evaluation pipeline, with configs, seeds, logs and every result, including negative ones.
Why fund this
- Concrete and measurable: standard public benchmarks (NorEval, EuroEval, BFCL) plus a new Norwegian tool-calling test set.
- Cheap to start, compute-bound to finish: the pilot runs on our own 16 GB GPU. The grant pays for the full-data runs and ablations a single card can't do in reasonable time.
- Open by default: Apache-2.0 base models, permissively licensed data, and open weights and code.
- Real use: local Norwegian assistants that can actually take actions (calendar, forms, search), without cloud lock-in.
Status (honest)
| Item | Status |
|---|---|
| Project plan, data and benchmark selection | DONE (this site) |
| NO-FC v0 benchmark + evaluator + pilot scripts | DONE (majere-nofc) |
| Pilot on a single 16 GB GPU (QLoRA, 7–8B) | NOT STARTED |
| Scaled training run (needs grant compute) | NOT STARTED |
| First baseline (no fine-tuning) | DONE qwen3:8b: 33/40 exact on NO-FC v0. See Data & evaluation |
| Fine-tuned results | NONE YET |