Phocinae-Largha-150M: a 144M-parameter model handles 82% of agent decisions

A developer released Phocinae-Largha-150M-v1, an open 144.3M-parameter typed decision model that answers agent questions (yes/no, pick-one, score) in a single forward pass. It runs in 18.6 ms on GPU and about 1.5 s on CPU; with a 0.6 confidence gate, LLM calls drop 82% while combined accuracy rises from 0.789 to 0.7948.

• 144.3M-parameter model returns a verdict with calibrated confidence in one forward pass
• GPU: 18.6 ms per decision; CPU-only: about 1.5 s; Apache-2.0 license
• English typed-decisions accuracy: 0.797 (400 cases, 2,000 decisions)
• A 0.6 confidence gate cuts LLM calls 82% and lifts accuracy to 0.7948