Jev directory
CommunityJev Models & Reimplementations
Jev-Style-2B-Decision-v4-MLX
MLX builds of Jev-Style-2B-Decision-v4 for Apple silicon: bf16 (3.76 GB) and 8-bit (2.00 GB) folders for text and JSON states, run through one shared jev_style_decision_mlx.py runtime under Apache-2.0.
- Reading time
- 2 min
- Published by
- Community
- Added
- Oct 2026
Highlights
- Text-classification model repo with bf16 (3.76 GB) and 8-bit (2.00 GB) weight folders in one repository.
- One runtime, jev_style_decision_mlx.py, selects a folder with --model-dir bf16 or --model-dir 8bit.
- Context per call is 32,768 tokens; on JevBench v1.4.1 public items the model scored 74.9%.
- 8-bit matched the FP32 reference top-1 on 99.9% of 1,000 gate questions; bf16 matched on 99.7%.
- On an Apple M1 Max, a bf16 question over up to 1,024 tokens took a median 233 ms.
Questions about this entry
Short answers drawn from the linked project or page.
Install huggingface_hub with hf_xet and run hf download chaoliangUNSW/Jev-Style-2B-Decision-v4-MLX --local-dir, or install jev-style with the mlx extra and load it via JevStyle.from_pretrained.
The bf16 folder holds 3.76 GB of weights and stays closer to the FP32 reference (max abs delta p 0.036); the 8bit folder holds 2.00 GB and made the same top-1 call as FP32 on 99.9% of 1,000 gate questions.
Context per call is 32,768 tokens, up from 25,600 in v3. A long fixture includes development rows from 4,097 up to 32,768 tokens, including an input of exactly 32,768 tokens.
The MLX builds were not re-run on the benchmarks. They reproduce the PyTorch runtime's scores within pre-declared gates, and the published Decision Index and JevBench numbers are neither official board entries.
The build is intended for on-device use on Apple silicon; the repository documents local execution with MLX and does not describe sending state or questions to a hosted service.
The repository is released under the Apache-2.0 licence, and it also ships LICENSE, NOTICE and THIRD_PARTY_NOTICES.md files alongside the weights.
Image states need the main repository's PyTorch runtime, and this build targets Apple silicon through MLX, so other hardware or multimodal work should use the main repository instead.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
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 NewsMore Jev resources like this
Russian-first open typed-decision model on Hugging Face. Given a context and a set of candidate options it returns a calibrated probability and picks one, covering yes/no, choice, and ordinal scale decisions. Built on mmBERT-base with 307M parameters under an MIT license.
Jev Models & Reimplementations#open-models#calibration#classification
Community
A 561M-parameter XLM-RoBERTa-large decision model given one extra training epoch on Jev's answers, with about 23% of examples black-box distilled from Jev's public API. Supports noul, choice, and score questions in 18 languages and loads through the bundled naluri.py.
Jev Models & Reimplementations#open-models#calibration#benchmark
Community
A trained example bank for @ruvector/typesafe on Hugging Face, covering 150 CLINC150 intents across 10 domains plus 1,000 out-of-scope utterances. It stores labelled examples and frozen splits, not weights; heads and temperature calibration are refit on load.
Jev Models & Reimplementations#open-models#classification#local
Community
A 46M-parameter Apache-2.0 decision model on Hugging Face, fine-tuned from q-project/N-Q-26M-JEV for typed classification. It exposes the custom NQ4ForDecision class to Transformers and lists Russian and English, plus NLI, sentiment, news, and toxicity datasets.
Jev Models & Reimplementations#open-models#classification#calibration
Community
Hugging Face text-classification checkpoint from Sagicc: a 4.9M-parameter F32 safetensors model with custom code, MIT licensed, documented with JEV decision APIs through AutoModel and a FastAPI server.
Jev Models & Reimplementations#open-models#python#local
Community
A 26M-class decision model on Hugging Face, fine-tuned from q-project/N-Q-26M-Chat. It ships a custom NQ4ForDecision class for text classification and feature extraction, is tagged calibrated and decision-model, and publishes Apache-2.0 weights with Russian and English support.
Jev Models & Reimplementations#open-models#calibration#classification
Community
From 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.