Jupiter.

    Our Sovereign Frontier Model Series.

    The Jupiter series

    Post-trained frontier models you can own, adapt, and deploy on your terms.

    Jupiter-N

    Our sovereign post-training of NVIDIA's Nemotron-3-Super-120B, adapted for British context with stronger instruction following, agentic behaviour, safety, and Welsh language capability, without regressing the base.

    Jupiter-G

    The same Locai post-training recipe applied to Google's Gemma-4 base, lifting instruction following, safety and code while preserving general knowledge.

    Frontier Performance.

    Request API Access

    Jupiter-N

    Ours

    vs. Nemotron base

    Jupiter-N
    Nemotron base
    AgenticTerminal Bench 2 (medium)
    ↑ +9.1 vs Nemotron-3-Super-120B
    Jupiter-N52.70
    Nemotron-3-Super-120B43.60
    Instruction FollowingIFBench (reasoning on)
    ↑ +4.1 vs Nemotron-3-Super-120B
    Jupiter-N73.80
    Nemotron-3-Super-120B69.70
    SafetyAgentHarm (lower is better)
    ↑ -1.6 vs Nemotron-3-Super-120B
    Jupiter-N53.80
    Nemotron-3-Super-120B55.40

    All values in %. Reasoning on/off configurations match those reported in the Jupiter-N technical report. Welsh benchmarks evaluated reasoning-off.

    Jupiter-G

    Ours

    vs. Gemma-4 base

    Jupiter-G
    Gemma-4 base
    CodeLiveCodeBench v6
    ↑ +3.2 vs Gemma-4-E4B-it
    Jupiter-G55.20
    Gemma-4-E4B-it52
    Instruction FollowingIFEval (prompt strict)
    ↑ +1.7 vs Gemma-4-E4B-it
    Jupiter-G89.30
    Gemma-4-E4B-it87.60
    SafetyAgentHarm harm rate (lower is better)
    ↑ -10.3 vs Gemma-4-E4B-it
    Jupiter-G12
    Gemma-4-E4B-it22.30

    All values in %. Same Locai post-training recipe (Forget-Me-Not™ replay + curated SFT) applied to a different open base. General knowledge held within ~1.4 pts of base while gains stack on instruction following, safety and code.

    Go deeper into the science

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