Jev directory
CommunityJev Models & Reimplementations
Decider 2B Vision
Mapika's vision decider: decider-2b v5 text weights transplanted into the Qwen3.5-2B vision-language model, tuned so an image plus a lettered question returns a calibrated probability at one answer slot with no generation. Breakout from pixels scores 41, above its teacher's 22.
- Reading time
- 2 min
- Published by
- Community
- Author
- Mapika
- Last verified
- Oct 2026
Highlights
- decider-2b v5 weights transplanted into the Qwen3.5-2B vision-language model, fine-tuned on images plus questions.
- Returns one calibrated probability distribution over lettered options at a single answer slot; no text is generated.
- Held-out Visual7W 0.89 with ECE 0.03; A-OKVQA 0.85, AI2D 0.93, ScienceQA 0.95, IconQA 0.94, 300 items per task.
- Play from pixels: Breakout 41 against the RAM-state teacher's 22, Pong 3 against 8, CliffWalking -13 (optimal).
- Apache-2.0; a 256x240 game frame costs 64 visual tokens, and text-only questions fall back to v5 behaviour.
Quickstart
python
import torch
from decider.vision import VisionDecisionModel
from decider.infer import Example, Q
m = VisionDecisionModel("<this repo>", grad_ckpt=False).cuda().eval()
ex = Example("This is a visual question about the image.",
[Q("What is the person holding?", ["a phone", "a cup", "a book", "nothing"], 0)])
inp = m.prepare([(image, ex)])
probs = torch.softmax(m.slot_logits(inp), -1)[0, :4]Watch out
Card reviewed 2026-10-17; every number is self-reported on the card, the text weights inside are decider-2b v5 so text-only behaviour lacks the v6 to v10 changes, and game play is measured over three episodes per game. Calibration is measured on the listed datasets, not on your images.
Questions about this entry
Short answers drawn from the linked project or page.
Mapika's vision variant of decider-2b: an image (a photo, a diagram, a game frame) plus a text question with lettered options yields a calibrated probability over the options at one answer slot, with no generation.
It keeps the no-generation property of System One models: options in, probabilities out. The family shares one interface, decider.infer.Decider and POST /v1/systemone in TypeSafe's format, though the vision path is used through the Python VisionDecisionModel API.
A 256x240 game frame costs 64 visual tokens. prepare takes a PIL image, numpy array or PNG bytes, and None for text-only questions, which run with the behaviour of decider-2b v5 including its abstention handling.
Pong and Breakout frames 0.96 with ECE 0.02, held-out Visual7W 0.89 with ECE 0.03, and multiple-choice image tasks between 0.80 and 0.95. 300 items per task, measured by the authors.
Breakout 41 (teacher 22), Pong 3 (teacher 8), CliffWalking -13 (optimal) and MiniGrid Empty 0.96; held-out Freeway, FrozenLake and harder grid worlds score 0. Three episodes each, so variance is high.
One epoch over 80k examples, 50k with images: game frames from Pong, Breakout, CliffWalking, MiniGrid and Mario labelled by scripted policies, plus multiple-choice image tasks from The Cauldron and a text replay, then PPO from pixels on Breakout and Pong.
It is not a chat model or captioner and only answers lettered options at one slot; the embedded text weights are decider-2b v5, so newer v6 to v10 input shapes and calibration are missing, and it is English only with calibration measured on the listed datasets.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsX: Happy JevDay: images, box, longer context, and an opt-in to find jaggedness
X: Standard One SH (StandardThinking)
More Jev resources like this
An Apache-2.0 decision engine with a TypeSafe-compatible endpoint, document answers with source passages, and a ViZDoom demo that picks actions from structured game state.
Jev Models & Reimplementations#community#python#local
Community
deepanwadhwa/OpenDecisionPython
deepanwadhwa/OpenDecisionPythonA local System One-compatible decision model on Qwen2.5-0.5B-Instruct: 50.7% accuracy and 0.245 ECE on JevBench public with a 13.1 ms median, built for demos rather than production.
Jev Models & Reimplementations#community#open-models#local
Community
A local yes/no decision model that speaks Jev's wire format on a laptop CPU, refusing Choice and Score until those question types ship.
Jev Models & Reimplementations#community#python#open-models
Community
feder-cr/jevC++
feder-cr/jevC++A local server that reads next-token probabilities from a fine-tuned Spark model and returns Choice, Score, and Noul answers with zero generated tokens.
Jev Models & Reimplementations#community#python#open-models
Community
Rizzo-AI-Academy/rizzo-flowPython
Rizzo-AI-Academy/rizzo-flowPythonA local /v1/systemone server that reads label probabilities from a GGUF chat model, with prompt repetition and a confidence shrink toward uniform.
Jev Models & Reimplementations#community#python#open-models
Community
Mushroom-Systems/lichenPython
Mushroom-Systems/lichenPythonJev-compatible decision models with vision input: LoRA fine-tunes of Qwen3.5 at 4B and 9B serving Choice, Score and Noul on the /v1/systemone wire format, with screenshots and photos placed beside text in the state. Runs on your own NVIDIA GPU; weights are non-commercial.
Jev Models & Reimplementations#community#vision#open-weights
Community
Xiaooolong/vevPython
Xiaooolong/vevPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.