Organization repository on GitHub
Phocinae decision-model inference server: TypeSafe /v1/systemone-compatible local API (Largha 150M)
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
Jev directory
CommunityJev Repos & SDKs
Phocinae Server
A local FastAPI inference server for the 150M-parameter Phocinae-Largha decision model that speaks the TypeSafe /v1/systemone contract: one POST returns noul, choice and score answers from a single forward pass, serving only 127.0.0.1 and never implementing chat or OpenAI-compatible endpoints.
- Stars
- 0
- Language
- Python
- License
- APACHE-2.0
- Reading time
- 2 min
- Published by
- Community
- Author
- Phocinae
- Last verified
- Oct 2026
Used inJev for Software & DevTools
Highlights
- Endpoints: /v1/systemone plus /permute for choice option-order averaging and /batch for up to 64 requests; /health needs no auth.
- Up to 64 questions per request, 255 choice options, an 8192-token context budget and a 2 MiB body cap (over-limit requests get 413).
- A 0.1.6 rendering bug left out-of-the-box accuracy at 0.5140 over 2,000 decisions; version 0.1.7 fixed the question rendering and reports 0.9055.
- Pure torch forward pass with no extra runtime frameworks; weights read from the published hf_repo safetensors and logits aligned with the official stack.
- Published latency: GPU fp16 p50 21.0 ms on an RTX 5090, CPU single-thread 1.64 s; the service measured ~36 ms p50 per CPU decision on 8 threads.
Quickstart
bash
git clone https://github.com/Phocinae/phocinae-server.git
cd phocinae-server
python -m venv .venv && . .venv/bin/activate
pip install fastapi uvicorn torch
PHOC_MODEL_DIR=/path/to/hf_repo python -m phocinae.mainWatch out
Grounded in the Chinese-language README and GitHub metadata reviewed on 2026-10-15; the latency and accuracy figures come from the author's RTX 5090 and 20-core CPU, the model weights are a separate download, and the README warns the service is not a complete security control and should fail closed.
Questions about this entry
Short answers drawn from the linked project or page.
A local inference server for the 150M-parameter Phocinae-Largha decision model. It listens on 127.0.0.1 and serves a TypeSafe System One-compatible API so agent tools can call POST /v1/systemone without leaving the machine.
No. It is a separate 144.3M-parameter model that implements the same request contract; the README describes it as a local decision service, not a clone of Jev's weights. Accuracy figures are the model's own.
Clone the repo, create a virtual environment, pip install fastapi uvicorn torch, then point PHOC_MODEL_DIR at the hf_repo directory and run python -m phocinae.main. It defaults to http://127.0.0.1:8155 with interactive docs at /docs.
No, deliberately. The README says a decision model is not a chat model and that chat, Ollama and OpenAI-compatible endpoints are intentionally not implemented; it answers noul, choice and score questions only.
Published figures: GPU fp16 p50 of 21.0 ms on an RTX 5090 and 1.64 s single-threaded on CPU. The running service measured about 36 ms p50 per decision on CPU fp32 with 8 threads; your hardware will differ.
The README warns against it: the server binds 127.0.0.1 with no auth by default, PHOC_BEARER_TOKEN can add a bearer requirement, and it advises fail-closed callers and a deterministic rule layer around any approval gate.
A default question-rendering bug in 0.1.6 left out-of-the-box accuracy at 0.5140 over 2,000 decisions. Version 0.1.7 uses each question's own instructions and criteria and reports 0.9055 over the same 2,000.
More Jev resources like this
A Jev-kind model inference engine and compatible API server in Python: each candidate path runs as segmented forwards with prefix reuse, and the Jev wire contract is served on top with NanoJev as the wired backend. The README reports 2.57x single-request throughput on MPS at 100% argmax agreement.
Jev Repos & SDKs#community#catalog-import#logicrw
Community
zerodegress/jevinfPython
zerodegress/jevinfPythonSelf-hosted Python SDK and System One-compatible API for runtime-defined Choice, Boolean and Score decisions on pinned open models, across Torch, vLLM, MLX, llama.cpp and WebGPU. Each result records probabilities, token usage, timing, model revision and prompt version, with committed row-level benchmarks.
Jev Repos & SDKs#community#sdk#self-hosted
Community
chengyongru/fastjevPython
chengyongru/fastjevPythonA native Swift implementation of OpenJev, the Jev-compatible typed-decision server: send a state and typed questions over a Jev-compatible HTTP API and read every answer from model probabilities, on Apple silicon Macs or inside iPhone and Mac apps.
Jev Repos & SDKs#community#catalog-import#cobanov
Community
algorythm-canada/openjevswiftSwift
algorythm-canada/openjevswiftSwiftA frozen decision server that serves Gemma 4 31B — quantized to FP8 at startup — in TypeSafe's /v1/systemone format, reading the next-token distribution instead of generating tokens. It needs one FP8 NVIDIA GPU with 80 GB and has no authentication.
Jev Repos & SDKs#community#github#python
Community
krishna-gogineni-765/deck31bPython
krishna-gogineni-765/deck31bPythonOpen reimplementation of the Jev decision-model idea: masked-logit softmax turns any small local LLM or OpenAI-compatible API into a type-safe decision engine with raw probabilities. Also ships a WeChat/QQ/Telegram crush radar with a verdict and next-move advice.
Jev Repos & SDKs#community#catalog-import#logicrw
Community
lookski/openjevPython
lookski/openjevPythonA lightweight Python engine that mimics Jev's structured decisions on local LLMs via Ollama, vLLM or llama.cpp: choice, score and noul questions come back as JSON with probabilities and confidence, plus an OpenAI-compatible gateway.
Jev Repos & SDKs#community#catalog-import#logicrw
Community
tapsin/jev-localPython
tapsin/jev-localPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.