Jupiter-N-120B fine-tuned tested
A 120B model post-trained on Welsh, with public weights, methods and datasets.
The facts
| Developer | Locai Labs / UCL (technical report: George Drayson et al., arXiv:2604.17429, 19 April 2026) |
|---|---|
| Base model | NVIDIA Nemotron-3-Super-120B-A12B — hybrid Mamba-2/MoE/attention LatentMoE, a one-million-token context window, and switchable reasoning |
| Method | Post-training through LoRA (r=16, α=32), with the ‘Forget-Me-Not’ anti-forgetting framework replaying synthetic data from the base model |
| Licence | NVIDIA Nemotron Open Model License; weights on Hugging Face |
Welsh: measurable data
Improving Welsh was one of the project’s three stated goals, alongside agentic ability and alignment with Britain’s cultural context. It used parallel corpora from Bangor University’s work — Senedd proceedings and UK legislation — alongside instruction data translated by a language model. According to the technical report, the result was:
An increase of 18 points on Welsh ARC-Easy and 5.25 points on Welsh MMLU-Lite compared with the base model, without losing general capability, thanks to the replayed synthetic data.
The datasets were published alongside the model weights: nemotron-chat-welsh, wikimedia_welsh, opensubtitles_welsh, eubookshop_welsh, legislation-gov-uk-en-cy and cofnodycynulliad_en_cy, the Senedd record. This lets others inspect and reuse the project’s data and methods. The paper describes the work as a ‘reproducible template for sovereign post-training’.
CymraegBench result: all 33,657 cases
The model was tested with BF16 weights on the full public suite, not just a sample. It completed 33,534 cases; 123 (0.37%) were recorded as invalid, with no infrastructure errors remaining at the end of testing. Jupiter scored 48.87 and placed eighteenth on the 23-model leaderboard; Mwydryn 7B v2, the highest-scoring open Welsh 7B model in the table, scored 30.37.
| Score | Result |
|---|---|
| Overall score | 48.87 |
| Welsh reasoning | 52.70 |
| Practical Welsh | 45.04 |
| Legislation translation — BLEU | 43.85 |
In the CCC Language Audit, Jupiter scored 89.17 on Welsh instruction following, 95.83 in English, 90.00 on summarisation and quotation, and 75.00 on idiom. The lowest scores were correction and explanation (17.03), vocabulary (7.53) and detecting offensive words (0.87). The scores by suite, CCC dimensions and metadata are openly available.
Bangor AI’s experiments with the model
Jupiter-N is the model Bangor AI has chosen for further fine-tuning on strict-metre poetry. On a cluster of eight DGX Spark machines, the team is running LoRA experiments on cynghanedd data to add traditional Welsh poetry to its experience of official and parliamentary Welsh. The full story, including what has worked and unresolved problems, is on the Jupiter experiments page.
What comes next?
- The cynghanedd protocol on the base model and our fine-tuned versions — before and after
- Repeating the test with reasoning enabled to measure the benefit and cost