Phocinae/Phocinae-Largha-150M-v1

What it solves

This model provides a fast, local, and predictable way to make structured decisions—such as approving commands, selecting tools, or filtering outputs—without the latency, cost, or privacy concerns of sending data to a large language model (LLM) API.

How it works

It is a 144.3M-parameter "typed decision model" based on the mmBERT-small architecture. Unlike a chatbot, it does not generate text. Instead, you provide a state (context) and a set of typed questions (Yes/No, multiple-choice, or 2–10 scoring). It performs a single forward pass to return a precise answer and a confidence score. It is designed to run locally on your own hardware, including CPUs, and outputs zero tokens.

Who it’s for

It is for developers building automated workflows, approval gates, or routing systems who need deterministic, low-latency decision-making. It is ideal for scenarios where decisions are repetitive and follow a fixed structure, but it is not suitable for open-ended conversation or long-form writing.

Highlights

• Zero-token output: Returns structured data (booleans, indices, or integers) directly.
• Local-first: Runs entirely on your machine; no API calls, no rate limits, and no data leaves your environment.
• High performance: Capable of 8–20 decisions per second on a standard CPU, or ~21ms per decision on a GPU.
• Bilingual support: Trained on both English and Chinese datasets.
• Decision routing: Can be used as a cost-saving gate, where only uncertain decisions are escalated to a larger, more expensive LLM.