Juno-N-Coder-25B-A3B-FP8

Juno-N-Coder-25B

NB: This is an FP8 quantised version of Juno-N-Coder-25B

Juno-N-Coder-25B is a coding-specialised derivative of NVIDIA Nemotron 3.5 Lightning 30B, produced by structured expert pruning with our SPACE algorithm (Specialisation Pruning Algorithm for Compression of Experts). This is the first model in the Juno series developed to run on-prem on the Locai One.

Locai Labs was one of NVIDIA's early-access partners for Nemotron 3.5 Lightning, and we want to thank the NVIDIA team for giving us early access to the model and supporting our work.

The goal with Juno-N-Coder was to create a coding-specialised version of Nemotron 3.5 Lightning purely through pruning and without any recovery fine-tuning. Using SPACE, we evaluated each expert's contribution to the target capability, in this case software development, removing the bottom 25% and renormalising the router so that the gate distribution remains correctly scaled.

The results are summarised in the model's performance profile below. Juno-N retains performance across four software engineering benchmarks, while degrading in general knowledge and factual recall as measured by MMLU Redux and SimpleQA.

All models were evaluated using NVIDIA NeMo Evaluator at identical settings for both models: temperature 1.0, top_p 0.95, a 65,536-token generation limit, BF16 weights, and reasoning traces separated from the response before scoring.

Usage

vllm serve locailabs/Juno-N-Coder-25B \
    --trust-remote-code \
    --max-model-len 131072 \
    --reasoning-parser ultra_v3 \
    --reasoning-parser-plugin "$PARSER" \
    --tool-call-parser qwen3_coder \
    --enable-auto-tool-choice \
    --enable-prefix-caching

Citation

@misc{juno-n-coder-25b,
  title  = {Juno-N-Coder-25B: Specialisation Pruning for Compression of Experts},
  author = {Locai Labs},
  year   = {2026},
  url    = {https://huggingface.co/locailabs/Juno-N-Coder-25B}
}