Jupiter.

    Our UK Sovereign Model Series.

    Explore modelsNeed your own model?

    Jupiter Series

    Jupiter-N

    April 2026

    Post-train of NVIDIA Nemotron 3 Super 120B for British context, Welsh language, UK cultural grounding and stronger agentic performance. Forget-Me-Not keeps the base capabilities intact.

    Technical ReportHugging FaceJupiter-N

    Jupiter-G

    April 2026

    The same Locai post-training recipe on Google Gemma-4, lifting instruction following, safety and code without regressing general knowledge.

    Hugging FaceJupiter-G

    Benchmarks

    +9.1 pts

    Terminal Bench · Jupiter-N vs base

    +18 pts

    Welsh ARC-Easy · Jupiter-N vs base

    −46%

    AgentHarm · Jupiter-G vs base

    Jupiter-N · agentic, instruction & safety

    BenchmarkNemotron-3-Super-120BJupiter-N-120B
    Terminal Bench 2 (medium)
    43.60
    52.70
    IFBench (reasoning on)
    69.70
    73.80
    IFEval (reasoning on, prompt strict)
    90.20
    90.20
    AgentHarm (reasoning on, lower is better)
    55.40
    53.80

    Jupiter-N · Welsh & retained capability

    Welsh ARC-Easy
    54.00
    72.00
    Welsh MMLU-Lite
    56.00
    61.25
    GSM8K (reasoning on)
    93.56
    94.01

    Jupiter-G · sovereign post-training on Gemma

    BenchmarkGemma-4-E4B-itJupiter-G-8B
    LiveCodeBench v6 (pass@1)
    52.00
    55.20
    IFEval (prompt strict)
    87.60
    89.30
    IFBench (prompt strict)
    34.40
    35.40
    AgentHarm harm rate (lower is better)
    22.30
    12.00
    MMLU Redux
    83.40
    82.00

    All values in %. Figures match the Hugging Face model cards and Jupiter technical reports. Jupiter-N Welsh benchmarks are reasoning-off; Terminal Bench, IFBench, IFEval, AgentHarm and GSM8K use the reasoning-on configuration where reported. Gains are measured against each model's open base.

    Go deeper into the science

    Read our research focus on continual learning, our patents, and the full technical story behind the Jupiter family.

    View research