Decima
122M multilingual local decision model answering Choice, Noul and Score in milliseconds on CPU behind a /v1/systemone server.
Overview
Decima is a small local judge for LLM systems. Given a situation, a question and options, it returns probabilities in roughly 20 to 50 ms on one CPU core using int8 ONNX, with no GPU, no API key and no data leaving the machine. It can also run entirely in the browser via a Hugging Face playground. The model pairs a multilingual-e5-small backbone with a late-interaction option scorer, so options are scored independently and shuffling them never changes the answer.
It speaks TypeSafe's /v1/systemone format, so Jev code works after a one-line change, and ships presets for triage, guardrails, moderation, routing, PII, agent gating and knowledge-graph judging. The README describes it as calibrated and evaluated in 20 languages. According to a list summary, it scores 0.616 on the jabr classifier-benchmark v2, well below hosted Jev, so it trades accuracy for locality and speed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team? | billing · technical · sales |
| Choice | gate: should this agent command be allowed or denied? | allow · deny |
| Yes / no | read_only: is this command read-only? | yes · no |
How it works
- pip install decima-ai (or use the browser playground)
- Pass text with a preset or custom question and choices
- Receive per-option probabilities in milliseconds
- Threshold, abstain or escalate to a larger model
Strengths
- Runs on one CPU core with no network
- Order-invariant option scoring
- Multilingual
- Jev-compatible API format
Caveats
- Lower accuracy than hosted Jev on a public benchmark
- Small model limits complex reasoning
Getting started
pip install decima-ai, then decima "I was charged twice" --question "Which team?" --choices billing,technical,sales.At a glance
- Type
- Benchmark
- Status
- beta
- Language
- Python
- Runtime
- CLI, library, server, browser (ONNX)
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Apache-2.0 122M multilingual decision model (multilingual-e5-small with a late-interaction option scorer) that answers Choice, Noul and Score questions behind a `/v1/systemone`-compatible server, in int8 ONNX on one CPU core (~20 ms for four options) or entirely in the browser ([playground](https://huggingface.co/spaces/amyrmahdy/decima-playground)); options are scored independently, so reordering them never changes the answer. It scores 0.616 on [jabr/classifier-benchmark](https://github.com/jabr/classifier-benchmark) v2, well below hosted Jev on the same benchmark.
cobanov/awesome-jev
Frequently asked questions
Does Decima need a GPU?
No, it runs int8 ONNX on one CPU core or in the browser.
Can I swap it for Jev?
It speaks the /v1/systemone format, so Jev code works after a one-line change, though accuracy is lower.
Does option order matter?
No, options are scored independently so reordering never changes the answer.
Related projects
- Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.
ServicebetaTypeScriptllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++Three open 4B decision models for English, Turkish and German that serve the Jev /v1/systemone API on a CPU in about 3 GB of RAM.
ServicebetaPython