Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

Mica

HTTP server for Mica, a 4B open decision model with 32k context serving yes/no, choice and score in TypeSafe's /v1/systemone format.

ServicebetaApache-2.0
View on GitHub ↗

Overview

Mica is a 4B open decision model for making decisions inside software: you send the material a decision depends on, the questions and the allowed answers, and it returns a probability for every answer with no text generated, so the cost of a decision is one read of the input. This repository is the HTTP server for Mica v0.5 4B Preview, running GGUF weights through llama.cpp and speaking TypeSafe's /v1/systemone format.

It supports up to 32,768 input tokens, 2 to 255 options or score levels, the three question types, and English, Korean and Chinese as main languages, under Apache-2.0. The README's example asks whether a full refund should be approved automatically and how urgent a ticket is, returning a yes probability and an expected urgency level with distribution. A Docker image builds for common CUDA architectures and downloads the Q8 weights on first start; it can also run without Docker on Linux with CUDA 12. Benchmarks, file sizes and limits are on the Hugging Face model card.

What the decision model decides

TypeQuestionAnswers
Yes / noShould the full refund be approved automatically?—
ScoreHow urgent is this ticket?Can wait · Today · Within the hour
ChoiceWhich of 2-255 options fits the input?—

How it works

  1. Build and run the Docker image (or the Linux setup)
  2. Weights download on first start
  3. POST state and questions to /v1/systemone
  4. Receive probabilities for every answer

Strengths

  • Long 32k-token inputs
  • Up to 255 options
  • Multilingual (English, Korean, Chinese)
  • Jev-compatible format

Caveats

  • Requires a CUDA GPU
  • Preview release
  • Benchmarks only on the model card

Getting started

git clone https://github.com/akivet/Mica && cd Mica && docker build -t mica . && docker run --gpus all -p 8010:8010 -v mica-weights:/weights mica

At a glance

Type
Service
Status
beta
Language
Python
Runtime
llama.cpp-based HTTP server (CUDA)
Has code
yes
License
Apache-2.0

Use cases and integrations

How curated lists describe it

  • A 4B decision model: calibrated probabilities for yes/no, choice and score questions over long inputs.

    hellogumbo/awesome-jev

  • HTTP server for Mica, a 4B open decision model (Apache-2.0, now v0.5 Preview) run as GGUF through llama.cpp: yes/no, a choice among 2 to 255 options, or a score, in TypeSafe's `/v1/systemone` shape. The Docker image needs a CUDA GPU.

    MrJev/awesome-jev

Frequently asked questions

How long can inputs be?

Up to 32,768 tokens.

Which languages does it target?

English, Korean and Chinese.

Do I need a GPU?

Yes, the Docker image needs a CUDA GPU.

Related projects

  • brier

    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
  • 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
  • browser-laya

    Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.

    ServicebetaTypeScript
  • ifReflex.cpp

    llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.

    ProjectbetaC++
  • jevalt

    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