CommunityJev Models & Reimplementations
Jev-Omni
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.
- Published by
- Community
- Author
- akhilaaa3
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- One classifier covers four modalities — text, image, audio and video — returning a probability per supplied option instead of generated text.
- Trained on Gemma 4 12B IT with a 30,000-question fine-tuning run; the card reports merged-model results.
- DecisionBench Medium 87.57% accuracy and JevBench 86.15%, with MMAU 63.10% and MVBench 53.10%.
- Warm H200 latency: 83 ms for about 2,000 text tokens, 26 ms image, 31 ms for 13-second audio and 504 ms for a 16-frame video.
- Apache-2.0 like Gemma 4; a CUDA GPU is required and FP32 weights need about 50 GB before runtime overhead.
Watch out
The 50 GB FP32 footprint needs a CUDA GPU, and while the head accepts 256 options the card says quality above 20 is not established. Benchmark figures are the author's own merged-model runs; dataset rights stay separate from the Apache-2.0 weights, and the card states it is not affiliated with TypeSafe AI.
Questions about this entry
Short answers drawn from the linked project or page.
Text, images, audio and video. You pass a state, a question and options, and add media plus modality 'image', 'audio' or 'video'. Audio is capped at 30 seconds and video uses 16 frames; the loader downloads the original Gemma 4 multimodal components.
On the card's table Jev-Omni's 63.10% MMAU trails Inkling's 77.20% (975B total, 41B active) and its 53.10% MVBench trails Qwen3.5-397B-A17B's 77.60%, though those reference scores come from their developers and may use different protocols.
Best supported at up to 20 options. The decision head accepts 256, but the card says quality above 20 is not established, so treat larger lists as unmeasured rather than supported.
A CUDA GPU. FP32 weights use about 50 GB before runtime overhead and inference runs under BF16 autocast. Warm H200 speeds are 83 ms for roughly 2,000 text tokens, 26 ms image, 31 ms for 13-second audio and 504 ms for a 16-frame video.
DecisionBench Medium 87.57% equal-weight scenario accuracy and 86.01% micro across 80 scenarios and 293 questions; JevBench 86.15% across 195 groups and 231 decisions; MMAU 63.10% on 1,000 questions; MVBench 53.10% on 2,786 questions.
Medium-set ECE is 0.0400 across 10 bins, lower being better, with the card's reliability graph drawn at five bins for readability. Because the classifier generates no output tokens, its cost comparison prices it at one call per question.
snapshot_download('akhilaaa3/Jev-Omni'), add the path to sys.path, then load_jev_omni() and call classifier.predict with state, question and options. Add media and modality for non-text inputs.
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 NewsMore 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
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.
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
Mithril's open, Jev-shaped typed-decision model on DeBERTa-v3-large: choice up to 255 options, score and noul from one forward pass, 0.854 in-domain with ECE 0.022, trained on public gold labels only.
Jev Models & Reimplementations#community#open-models#local
Community
Vision-capable GGUF port of autotrust/JEV-27B-VL: BF16 through Q3_K_M with a separate mmproj file, running the calibrated System 1 typed-decision interface and a 27B System 2 through llama.cpp.
Jev Models & Reimplementations#community#open-models#local
Community
An EmbeddingGemma 2 fine-tune that answers choice, score, and noul questions from a state and a plain-data schema, with a 0.5M-parameter schema-conditioned scorer over 768-d embeddings.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
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
Jev takes text only; Clef-omni adds image, audio, and video input and Liquid d1 adds images, each keeping the same typed-answer shape. Where that matters, where it does not, and the costs the vendors underplay.
Jev Models & Reimplementations#multimodal#open-models#decision-model
Read article
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
From the Jev community
Posts from builders shipping with Jev right now.