Phocinae-Largha-150M-v1（斑海豹 Largha）

Phocinae-Largha-150M-v1（斑海豹 Largha） is a 144.3M bilingual typed decision model that processes state and questions in a single forward pass with zero generated tokens.

Phocinae-Largha-150M-v1 is a 144.3M parameter bilingual typed decision model built on jhu-clsp/mmBERT-small. It is designed for repetitive and closed-option AI agent decisions such as command approval, tool selection, step validation, and output screening. Instead of generating conversational text, it takes input states and typed questions to output calibrated probability distributions in a single forward pass. It runs locally with or without a GPU, reducing LLM API calls significantly. The model features open-source code under the Apache-2.0 license.

Phocinae-Largha-150M-v1（斑海豹 Largha） at a glance

Best for: Software engineer
Runs on: self-hosted, open source
Pricing: Apache-2.0 open source
Released: October 2026
Category: AI agent tool
Website: phocinae.github.io

Key features

Run local typed decisions
Process state and typed questions
Return calibrated confidence scores
Execute single forward pass
Generate zero output tokens
Support CPU and GPU execution

Use it for

Approve commands in AI agents
Select tools for workflows
Validate execution steps
Screen model outputs

Alternatives to Phocinae-Largha-150M-v1（斑海豹 Largha）

OpenSquilla — Maximizes AI intelligence density while minimizing token costs through smart routing.
Composio — Equips AI agents with tools, managed user auth, and sandboxes across 1,500+ apps
Laya Mlx — Provides local MLX inference with no PyTorch or cloud API dependency.
Waku Agent — Local-first memory storage and readable Python code for AI agent logic.
Rufous — It runs locally, checks small model tool calls, and asks before making system changes.
Overview — Returns deterministic probabilities that sum to 1 with zero free text written.
WebBrain — Runs locally or in the cloud with your choice of LLM and zero telemetry.
Beeuni — Agents retain context and improve skills over time through a controlled review process.
Melra — Governs AI agent tool calls with local policies, approval gates, and verified receipts.
NemoClaw — Provides managed security, networking, and lifecycle operations for AI agents.
Klavis — Provides live environments for training agents with verifiable reward data.
Kodama — Provider-neutral multi-agent workflows for the Kiro CLI with persistent project memory.

All Phocinae-Largha-150M-v1（斑海豹 Largha） alternatives