vadery/netis-n1

Netis N1

Netis N1 is a 32B-parameter model specialized for intelligent IT operations (AIOps), fine-tuned from Qwen2.5-32B-Instruct using Continual Pre-training, SFT, and RL.

Highlights

• OpenRCA Scoring 26.0% — on par with Claude Opus 4.5 (26.9%), surpassing GPT Codex (21.0%) and DeepSeek (13.6%)
• General capabilities preserved or improved (GPQA +9.6%, MMLU-Pro +3.3%, IFEval +1.0%)
• Two-stage diagnostic workflow: tool-call planning + fault localization

Evaluation Results

Benchmark	Base (Qwen2.5-32B)	Netis N1
OpenRCA Bank (scoring)	—	26.0%
MMLU-Pro (300)	63.7%	67.0%
GPQA Diamond (198)	44.4%	54.0%
IFEval (100)	86.0% / 90.5%	87.0% / 91.2%
GSM8K (1319)	91.5%	91.4%
AIME (90)	16.7%	11.1%
HLE (100)	2.0%	1.0%

How to Evaluate

Serve with vLLM and run the eval suite:

# Start vLLM server
python -m vllm.entrypoints.openai.api_server \
  --model ./netis-n1 --port 6777 --max-model-len 20000

# OpenRCA eval (requires OpenRCA Bank dataset)
python eval_model.py --llm-backend vllm --vllm-base-url http://localhost:6777

# General benchmarks (MMLU-Pro, GPQA, AIME, IFEval)
python eval_multi_model.py --name netis-n1 --url http://localhost:6777 --model netis-n1

License

Apache 2.0