Phocinae-Largha-150M-v1 is now released, a 144.3M bilingual model that handles routine agent decisions locally with zero generation tokens.📜 Apache 2.0.
🤖 https://modelscope.ai/models/PerryLink/Phocinae-Largha-150M-v1
🎯 Handles yes/no, pick-one, and 2–10 scoring questions with calibrated confidence in a single forward pass.
🏆 Scores 0.906 on the fitted English typed-decisions evaluation and 0.848 on the in-mix Chinese evaluation.
⚡ Delivers 21 ms GPU latency in the reported RTX 5090 setup; its FP16 weights occupy just 288.6 MB and CPU inference is supported.
💰 At a confidence threshold of 0.6, its escalation router reduces LLM calls by 55% while retaining 0.9936 accuracy on the high-confidence subset.