Phocinae/Phocinae-Largha-150M-v1 — new model trending #28 on Hugging Face

Phocinae releases Largha-150M-v1, a 144.3M-parameter structured-decision model claiming 21 ms GPU latency and 55% fewer LLM calls.

Phocinae-Largha-150M-v1 is an Apache 2.0, 144.3M-parameter bilingual (English/Chinese) decision model built on an mmBERT-small base with 8192-token context, designed for one-forward-pass structured decisions rather than chat. It reports 0.906 typed-decision accuracy on English and 0.848 on machine-translated Chinese, with 21.0 ms p50 GPU latency on an RTX 5090. A confidence-based escalate gate (tau=0.6) is claimed to cut LLM API calls 55% while raising kept-subset accuracy to 0.9936, targeting agent routing, command approval, and tool selection tasks. It scored 0.5455 on JevBench public-231, below its own 58.4% gate, which the authors disclose.