Mapika/decider
A family of System One-style models fine-tuned from Qwen3.5, designed for one-pass typed decisions with calibrated probabilities.
What it solves
Traditional language models generate free-form text. When you need a decision among a fixed set of options (like routing a ticket, classifying intent, or answering yes/no), you typically have to parse the generated text, which is error-prone and gives poorly calibrated confidence. decider does away with text generation entirely: it reads a state and a set of typed questions, and in one forward pass returns a probability distribution for every question over the allowed options. No decoding, no parsing, no output outside the options you defined.
How it works
The model takes a state (the context or input) and a set of typed questions. Each question has a fixed answer set: Choice (2 to 255 options), Score (2 to 10 described levels), or Noul (probability of yes). A single forward pass produces a distribution for every question. The models are trained on a mixture of public datasets plus teacher-labeled data from a local Qwen3.5-27B model, and the project is an independent open reproduction of the "System One" model class (such as TypeSafe AI's Jev), with no distillation from Jev. The package provides a Decider class with a decide method, an HTTP server (decider.serve), GGUF loading for llama.cpp, and integration with vLLM for large-scale serving. Multiple model sizes (0.8B, 2B, 4B, 12B, 31B, 35B) and variants (some handling images) are released on Hugging Face, each with configurable per-type temperatures and optional checkpoints.
Who it's for
Developers and researchers who need fast, calibrated decisions from language models — for routing, classification, evaluation, game-playing (as shown in Tetris and Pong demos), or any application that wants a probability per option without generating text. The project runs on CPU, Apple Silicon (MPS), CUDA GPUs, llama.cpp (GGUF), and even a Snapdragon X Elite NPU via a community port, so it suits everything from laptops to data-center GPUs.
Highlights
- One-pass typed decisions with calibrated probabilities — no text generation, no parsing.
- Models from 0.8B to 35B parameters, including vision-language variants for image-based decisions.
- Strong leaderboard standings: decider-4b v2 is #7 on JevBench, and decider-chat-gemma4-31b is #2 on the Decision Index.
- Flexible deployment: CPU, MPS, CUDA, GGUF via llama.cpp, vLLM, and even NPU via community contributions.
- Independent open reproduction of the System One model class, with training code and data mixtures publicly available.
Related
- Dispatch
- Project
- Project
- Dispatch
- Dispatch