What is Phocinae-Largha-150M-v1?

A 144.3M-parameter bilingual typed decision model, not a chatbot: given a state and typed questions it returns calibrated answers with confidence in one forward pass, emits zero output tokens, and is insensitive to option order.

Which questions can it answer?

noul for yes/no with an optional threshold, choice for picking one of up to 255 options and returning a zero-based index, and score for a 2-10 rating; up to 64 questions can be asked in one request.

What hardware does it need?

No GPU is required. CPU single-thread p50 is 1.64 s per case and eight threads process 8-20 decisions per second; the card reports a GPU fp16 p50 of 21.0 ms on an RTX 5090, which it says was not independently reproduced.

How accurate is the model?

The curated English score is 0.791, with 0.989 on the 1,500 untouched cases and 0.198 on 500 contaminated ones; Chinese is 0.848 on an in-mix evaluation. The benchmark's teacher self-agreement ceiling is 0.735, and option-order flip rates are 0.0144 to 0.0283.

How do I run it locally?

Download the weights with hf download, install phocinae-server, and start it with PHOC_MODEL_DIR pointing at the weights; it serves HTTP on 127.0.0.1:8155 with a /v1/systemone endpoint, and the card shows a curl example.

Does it need a TypeSafe key or the Jev API?

No. It is a local model with its own runtime. Its protocol matches the System One shape with noul, choice, and score questions on /v1/systemone, but inference runs on your machine and needs no key, account, or network.

What license does it ship under?

Apache-2.0 for this model, with the jhu-clsp/mmBERT-small base under MIT. Weights are 288.6 MB fp16 with a sha256 the card states it verified against the 288,603,268-byte file, and a calib column exists but is off by default.

Jev directory

CommunityJev Models & Reimplementations

Phocinae-Largha-150M-v1

A 144.3M-parameter bilingual typed decision model fine-tuned from jhu-clsp/mmBERT-small: it answers noul, choice, and score questions in one forward pass with calibrated confidence and zero output tokens, and it runs on CPU.

See highlights Open github.com

Stars: 167
Language: Python
License: APACHE-2.0
Reading time: 3 min
Published by: Community
Author: phocinae
Last verified: Oct 2026

phocinae/phocinae-largha-150m-v1

Organization repository on GitHub

View on GitHub

斑海豹 — Phocinae-Largha-150M-v1: a 144.3M bilingual decision model. Tiny model. Big decisions. 小海豹，大决断。

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

Tagscommunity catalog-import cobanov yibie decision-model bilingual cpu-inference

Works forJev for Classification Jev for Scoring Jev for Decision gating

Used inJev for Software & DevTools

Languages: Python
Platforms: API CLI
Capabilities: Open sourceSelf-hostedRuns locally

Highlights

Protocol: noul (boolean with optional threshold), choice (0-based index, up to 255 options), and score (2-10), up to 64 questions per request.
Weights are 288.6 MB fp16 with a stated sha256; inference peaks near 1.6 GB VRAM or 1.8 GB RAM and runs without a GPU.
Factory max_len is 512 for the whole state-plus-questions sequence even though the encoder supports 8192 positions; the decision head window is 192.
The card leads with a curated re-score of 0.791 on typed-decisions en and discloses 0.989 on the 1,500 cases training never touched.
Not a chat model and not loadable through the usual transformers path; use the phocinae-server runtime, which does its own forward pass.

Quickstart

bash

hf download Phocinae/Phocinae-Largha-150M-v1 --local-dir ./largha
pip install phocinae-server
PHOC_MODEL_DIR=./largha python -m phocinae.main

Watch out

Model card reviewed 2026-10-15. The card states the Chinese 0.848 is an in-mix (fitted) evaluation, that its GPU latency was not independently reproduced, that JevBench public-231 scored 0.5455, and that factory ECE calibration is 0.2519 en / 0.1941 zh unless the optional calib column is enabled. An earlier 0.906 headline was restated as 0.791.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

Beyond text and boxes: multimodal decisions you can run yourself

Oct 16, 2026

Jev-Omni's merged classifier scores text, image, audio, and video options; VideoJev and Laya Vision show what open multimodal decision models accept, what they return, and which numbers are the authors' own.

Jev Models & Reimplementations#multimodal#open-models#vision

Read article

Jev 1.14: Images, Box, and the Danger Zone

Oct 15, 2026

Jev Day shipped a routine model bump and a deliberate widening of the API: images on both models, a box question type on 1.14 only, and 262k context, all behind danger-zone headers.

Jev Models & Reimplementations#release#vision#danger-zone

Read article

The Open Decision-Model Roundup

Oct 14, 2026

Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.

Jev Models & Reimplementations#open-models#training#local

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai

Phocinae-Largha-150M-v1

Highlights

Quickstart

Questions about this entry

What is Phocinae-Largha-150M-v1?

Which questions can it answer?

What hardware does it need?

How accurate is the model?

How do I run it locally?

Does it need a TypeSafe key or the Jev API?

What license does it ship under?

More Jev resources like this

From the Jev guides

From the Jev AI model community