CommunityJev Models & Reimplementations
Jev-Style Qwen3.5-2B GGUF
chaoliangUNSW's v1 Jev-Style decision model in GGUF from 1.3 GB: calibrated choice, bool and score answers on Qwen3.5-2B-Base, 82.3% on five held-out task families with ECE 0.017.
- Published by
- Community
- Author
- chaoliangUNSW
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Three files: Q4_K_M at 1.3 GB keeps 82.4% accuracy, Q8_0 at 2.1 GB matches bf16 on 99.4% of decisions, BF16 at 3.9 GB on 99.8%.
- 82.3% accuracy on 1,500 held-out examples against Qwen3.5-2B-Base's 65.9%, with ECE 0.017 against the base's 0.065.
- The calibration temperature is folded into the final RMSNorm weight, so nothing is applied at inference time.
- Trained on 22k typed decisions from SST-2, MNLI, AG News, BoolQ and SST-5 via LoRA rank 16 with a log-score loss.
- 77 ms per decision on an M1 Max under MLX bf16, the card says at no added cost over the unmodified base model.
Watch out
Requires the card's prompt format — plain chat messages produce meaningless continuations — and probabilities read through a server's top_logprobs cap at 20 options against 26 native. Weights are Apache-2.0 like the Qwen base, but AG News training data is research/non-commercial, and the model keeps base-model accuracy on unseen task types.
Questions about this entry
Short answers drawn from the linked project or page.
It does not write text: give it a state, a question and options, and it returns the decision with calibrated probabilities from a single token position. This v1 supports choice, bool and score primitives and cannot answer outside the declared options.
Q4_K_M at 1.3 GB is the smallest but keeps 82.4% accuracy and matches bf16 on 94.4% of decisions. Q8_0 at 2.1 GB matches on 99.4% and scores 81.6%, and BF16 at 3.9 GB matches on 99.8% and scores 81.6%.
No. You must use the prompt format ending in 'Answer:' — the next token is the option letter. Plain chat messages produce meaningless text continuation, and the card says to start a new chat for every decision if you paste the prompt manually.
ECE is 0.017 on 1,500 held-out examples against the base model's 0.065; the card says simulating labels from the model's own probabilities gives an expected ECE of 0.017 from sampling noise alone. A temperature was folded into the final RMSNorm weight so no post-processing is needed.
Up to 26 options natively, or 20 when probabilities are read through a server's top_logprobs. The client renormalises the option letters, so the returned distribution always covers the declared options.
SST-2 and MNLI from GLUE, AG News, BoolQ and SST-5, 22k examples converted to typed decisions, with 80% for LoRA training and 20% held out for calibration. Weights are Apache-2.0 like Qwen/Qwen3.5-2B-Base; AG News is research/non-commercial.
This repository preserves v1. The card's banner reports v3 at 0.8B scoring 79.2% on the 2,000 typed decisions against v1's 53.4% and v2's 73.5%, and describes v3 as taking 25,600-token inputs and 51 languages.
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: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsHacker 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 NewsRelated terms
Glossary definitions related to this entry.
More Jev resources like this
Reza2kn's llama.cpp Q4_K_M build of Jev-Omni: a 6.87 GiB multimodal decision backbone plus projector and FP32 head, measured at 8.98 GiB VRAM on an RTX 5080 Laptop GPU.
Jev Models & Reimplementations#community#open-models#local
Community
An Apple Silicon MLX bf16 build of the v1 Jev-Style decision model: 77 ms decisions on an M1 Max with calibrated choice, bool and score probabilities, Apache-2.0 weights.
Jev Models & Reimplementations#community#open-models#local
Community
AutoTrust's combined System 1 and System 2 bundle on Gemma-4-26B-A4B-it: Decision Index 0.2.1 of 58.05, a 24-slot head with noul, choice and score, and the base chat model left bit-identical.
Jev Models & Reimplementations#community#open-models#local
Community
akhilaaa3's merged multimodal decision classifier on Gemma 4 12B IT: probabilities for text, image, audio or video options, 87.57% on DecisionBench Medium and MMAU 63.10%, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
alibiserikbay's Apache-2.0 Jev alternative on Qwen3.5-4B with a merged rank-16 LoRA: one forward pass returns a probability per option, 0.784 on JevBench's public hard tier at about 13 ms per decision.
Jev Models & Reimplementations#community#open-models#local
Community
GGUF builds of the JevK5 family for llama.cpp: Q8_0 is effectively lossless (229 of 231 answers match the 4B v0.3 bf16), and every published file clears the card's 95% agreement gate.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.
Jev Models & Reimplementations#open-models#training#local
Read article
Interface, size, license, claims — in that order. A selection guide across the system-one catalog, from a 706K form scorer to 27B Jev students.
Jev Practices & Patterns#open-models#evaluation#benchmarks
Read article
From the Jev community
Posts from builders shipping with Jev right now.