WHAT-THE-JEV / RESOURCES
Decision models and the ecosystem taking shape around them.
Chinese fine-tune of Laya's multilingual checkpoint, with machine-translated data splits and author-reported test results.
Typed decision family from 0.8B to 35B-A3B reading text, JSON, and images, with non-commercial weights.
Bespoke Labs' open Jev alternative: a Qwen3.5-9B one-step typed text decision model with public data and recipe.
razorback16's open System One decision server; several unrelated projects share the OpenJev name.
Interfaze AI's System One decision model plus the levbench evaluation harness; adoption is early.
Supersonic Labs' 144.3M-parameter compact decision model that runs on CPU and in the browser.
Frontier Infra's System One-style decision model with a full training, data, and evaluation pipeline.
Invergent's multimodal decision model reading text and images and returning probabilities in one pass.
TokenRhythm's prefill-only decision model for agent routing, tool selection, and scoring.
Multilingual non-autoregressive decision engine returning typed decisions with calibrated confidence in one forward pass across 100+ languages.
Small Qwen-based decision model family for self-training and local deployment, with three isolated question types and calibrated probabilities.
Mapika's Apache-2.0 decision model family spanning 0.8B to 35B variants, with the training recipe published in full.
Open decision model pairing distilled Qwen3.5 LoRA with option-letter logit readout; ranked first among open-source models on JevBench v1.4.
Single-forward-pass decision server with five models; its 4B adapters are restricted to research and demonstration use.
Zero-training approach reading option probabilities from frozen open models in one forward pass, with MIT-licensed code.
Multimodal decision model fine-tuned from Qwen3.8-27B, with Apache-2.0 weights and a commercial hosted API.
Cloudflare decision models at 27B and 9B with image input, 64K context, and Jev API compatibility.
Open Qwen3.5 4B/0.8B LoRA decision weights supporting choice-type only, with a $17 training walkthrough.
KAIST-led open-weight judge model family outputting scores and written judgments, not typed calibrated probabilities.