go-system-one
Native Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.
Overview
Go System One uses Gemma 4 12B for finite-choice decisions, ordered scores and yes/no probabilities through a Jev-like API. The authors started from a fixed Gemma baseline, built a llama.cpp prototype to define expected results and speed, and then replaced it with a native Go runtime.
The NVIDIA path loads hand-tuned PTX kernels directly from Go through the NVIDIA Driver API, needing no CGo, llama.cpp or CUDA toolkit at run time. The README reports the pinned RTX 3060 request falling from about 521 ms to a 72.80 ms median, and automatic two-field batches reaching about 12 entries per second. On the JevBench v1.4.0 public subset it reports 84.85% (196 of 231) with no failed requests, noting this is public accuracy, not the official sealed composite score. A local web playground and batch decision interface are included.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Finite-choice decision over supplied options | — |
| Score | Ordered score over supplied levels | — |
| Yes / no | Yes/no probability question | yes · no |
How it works
- Download Gemma model files under its licence
- Run the Go server with the NVIDIA backend
- Send Noul, Choice or Score questions
- Read probabilities; use batch mode for many entries
Strengths
- No CGo or CUDA toolkit at runtime
- Low latency on a consumer GPU as reported
- Published JevBench evidence
Caveats
- Requires an NVIDIA GPU for the fast path
- Gemma licence applies to model files
Getting started
make setup, make artifacts-download ACCEPT_GEMMA_LICENSE=1 HF_TOKEN=..., make run BACKEND=nvidia LISTEN=127.0.0.1:8080.At a glance
- Type
- Project
- Status
- beta
- Language
- Go
- Runtime
- native server (NVIDIA Driver API)
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
When a gopher met Jev.
hellogumbo/awesome-jev
Frequently asked questions
Which model does it use?
Gemma 4 12B.
How accurate is it?
84.85% on the JevBench v1.4.0 public subset, per the README.
What hardware?
Benchmarks are on an RTX 3060; the native path targets NVIDIA GPUs.
Related projects
Go gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.
ServicebetaGoTurns any chat model on SGLang, vLLM, transformers or MLX into a Jev-compatible service with one prefill and one label token per question.
ProjectbetaPythonSelf-hosted Jev-compatible System One server that sits in front of llama-server and reads option probabilities from any GGUF model.
ServicebetaPythonInference engine for Jev-style decision models using segmented forwards with prefix reuse, serving the Jev wire contract.
ProjectresearchPythonResearch preview: Python client and pinned adapters for running System One decision models on your own Modal GPUs.
LibraryresearchPython- simple-jev★ 602
Featherless AI server that turns compatible open Hugging Face models into a /v1/systemone classifier by reading next-token logits.
ServicebetaPython