decider
Mapika open System One family on Qwen3.5: one forward pass, Choice/Score/Noul, POST /v1/systemone via decider-ai on PyPI, weights on Hugging Face.
Overview
decider (github.com/Mapika/decider, Apache-2.0, PyPI decider-ai import decider) is an independent System One class model family: state plus typed questions in, calibrated probabilities from one forward pass out, no text generation. Checkpoints include decider-0.8b through decider-35b-a3b and decider-2b-vision on Hugging Face under Mapika org tags. README states no affiliation with TypeSafe AI and no distillation from hosted Jev; teacher data comes from public mixture plus local Qwen3.5-27B labels. decider.serve and decider.serve_vllm expose TypeSafe wire-compatible HTTP; python -m decider.calibrate fits per-type temperature maps. README Standing section quotes JevBench and Decision Index leaderboard rows read on published dates; treat those as third-party harness results, not this directory's measurements. Showcase GIF on GitHub raw shows arcade games where each move is one forward pass.
Problem: Teams comparing open decision models still confuse Jared Palmer Kev, Von, Laya packaging, and independent Qwen trainers that all speak System One but ship different weights and eval stories.
Built for: ML engineers who want Mapika's Qwen3.5-based decider checkpoints, decider-ai serve on POST /v1/systemone, and public training mixture notes without TypeSafe API dependency.
First indexed on Jev Directory: 2026-09-27
Creator and team
- Name
- Mapika
- Handle
- @Mapika
How Jev is used
- Role in the product flow
- Self-hosted parallel Choice, Score, and Noul over shared state via native decider forward pass or /v1/systemone server
- Primitives
- ChoiceScoreNoul
- State in
- JSON state and questions map per README curl examples and decider-ai HTTP schema matching TypeSafe System One.
- Decision out
- Probability vector per question; no decoding step outside defined options.
- pip or uv install decider-ai and pull weights from Hugging Face
- python -m decider.serve or serve_vllm for HTTP clients
- POST /v1/systemone with parallel questions on one state blob
- Optional calibrate CLI on labelled rows before production gates
- Point TypeSafe SDKs at local base URL when you want drop-in clients
decider is its own model lineage, not a thin alias of Kev or Von. jaredpalmer-kev targets Jared's Qwen fine-tunes with SDK parity tables against hosted Jev; wfzyx-von chases non-autoregressive encoder speed; theoleecj-semif documents another open server narrative. ollaya-dev-ollaya can pull and serve decider ONNX wrappers for local :11435 workflows, but Mapika/decider and Hugging Face Mapika/decider-* repos remain training and weight source of truth. Compare realzachi-typesafe-adblock only when your question is browser clutter, not model training.
Sourced performance claims
- README JevBench table read 2026-09-21 lists decider-35b-a3b at 68.9 total score (#10 of 36) and decider-2b at 64.6 (#21 of 36).Source: github.com/Mapika/decider README Standing
- README Decision Index v0.1 dated 2026-09-22 ranks decider-35b-a3b NVFP4 fourth of 32 at 54.3 and decider-2b fourteenth at 44.0.Source: github.com/Mapika/decider README Standing
- Public GitHub repo Mapika/decider had 702 stars and 39 forks when this listing was drafted.Source: GitHub API September 2026
Features and stack
Features
- decider-ai PyPI package with decider.serve and serve_vllm paths
- Multiple HF checkpoints from 0.8B dense to 35B-A3b MoE
- Public mixture.py and teacher_data transparency in repo
- Game and browser RL stage notes with linked docs/DEMOS.md
- calibrate CLI for per-type temperature fitting
Stack
- Python
- PyTorch
- CUDA, Apple MPS, vLLM optional path
- Hugging Face weights
Pricing: Apache-2.0 code and published weights; you pay for GPUs and electricity. Third-party leaderboard hosting may bill separately.
Links
FAQ
- Is decider affiliated with TypeSafe?
- README states independent project, not affiliated with or endorsed by TypeSafe AI.
- Does ollaya replace this repo?
- ollaya packages ONNX serve for many open models including decider tags. Mapika/decider remains where training recipes and authoritative cards live.
- What PyPI package name?
- Install decider-ai; Python import remains decider per README.
- Can I use TypeSafe SDKs against decider.serve?
- README documents POST /v1/systemone with the TypeSafe wire format so compatible clients can point at localhost.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- System One model
The model family behind Jev: parallel typed questions, one forward pass, probabilities you can threshold.
- Jev vs LLM classification
When to gate with System One probabilities instead of asking a chat model to label things.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.