brier
Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
Overview
Brier is Mangaba AI's typed, calibrated decision model for Portuguese text. It runs locally on Windows, Linux or macOS on CPU, NVIDIA GPU or Apple Silicon, offline and with no per-call cost. Given content and one or more questions with predefined answers, it returns the chosen option and the probability of every option, never generating text. It is named after the Brier score, which its training optimizes so that probabilities match reality, not just answers.
The model is Qwen3-4B with a LoRA trained in Portuguese; v2 was retrained against instruction injection and the README says it resists injection better than Jev. Its own head-to-head comparisons report 72.0% versus Jev's 67.6% on tweetSentBR sentiment, better calibration than Jev on B2W product reviews with statistical significance, and a statistical tie on a 46-domain synthetic set. Versions from 5.0.0 are under a proprietary Mangaba AI licence; earlier releases (package 4.1.0 and v2 and v3 weights) remain Apache-2.0. Docs are in Portuguese with an English summary.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which predefined answer fits this Portuguese text (e.g. sentiment)? | — |
| Yes / no | Yes/no judgment about the content(inferred) | yes · no |
How it works
- Pass Portuguese content and questions
- Model scores every option locally
- Return chosen option and full distribution
- Apply thresholds in your code
Strengths
- Offline, no per-call cost
- Trained for calibration
- Reported better injection resistance than Jev
Caveats
- Proprietary licence from v5.0.0
- Comparisons are the author's own
Getting started
See plans and licensing at https://brier.mangaba.ia.br; Apache-2.0 versions are at the ultima-apache-2.0 tag.At a glance
- Type
- Project
- Status
- production
- Language
- Python
- Runtime
- local model (CPU/GPU/Apple Silicon)
- Has code
- yes
- Has tests
- yes
- License
- Proprietary (v5+); Apache-2.0 (up to 4.1.0)
Use cases and integrations
How curated lists describe it
Brier: typed, calibrated decisions (choice/score/noul) for Portuguese, running…
wh000wh000/awesome-jev-live
Frequently asked questions
Which language does it handle?
Portuguese text.
How does it compare to Jev?
Its own tests report higher sentiment accuracy, better calibration on product reviews and a tie on a synthetic set.
Is it open source?
Versions up to 4.1.0 are Apache-2.0; 5.0.0 and later are proprietary.
Related projects
- 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 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.
ServicebetaPythonJapanese System One decision model on ModernBERT-ja (314.6M) returning typed choice, score and bool distributions with no generation.
ProjectbetaPython