Jev directory
CommunityJev Models & Reimplementations
JEV-9B on ModelScope
AutoTrust's JEV-9B on ModelScope: a Blocks of Experts System 1 + System 2 model fine-tuned from Qwen3.5-9B, Apache-2.0, about 8.95 GB of BF16 safetensors in five shards, taking images since 3 October 2026, with a GuruSearch live demo.
- Reading time
- 3 min
- Published by
- Community
- Author
- AutoTrust
- Last verified
- Oct 2026
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Fine-tune of Qwen/Qwen3.5-9B with kl, auroc, brier, and ece in its metrics list and about 8.95 GB of BF16 and F32 safetensors across five shards.
- AutoTrust's first integrated System 1 + System 2 open model with the Blocks of Experts recipe: one set of weights and one vLLM engine, routed per request.
- The card says its System 1 decisions are indistinguishable from the closed TypeSafe Jev 1.13 by KL, with a decision in about a third of the hosted API's time.
- Takes images since 3 October 2026: camera image or screenshot in, a probability per action out in about 0.2 s on one GPU, including robot-arm pick and place.
- Trained on SargeDev/jev-distill-corpus-v3; the live GuruSearch demo, new 11 October 2026, ranks headlines from 6 news feeds and re-ranks search results.
Watch out
Apache-2.0 per the metadata. The KL-indistinguishability and speed claims are AutoTrust's own card statements, and ModelScope's metrics list names kl, auroc, brier, and ece without attached values, so no accuracy figures are visible in this capture. The GuruSearch demo and benchmark scripts are external evidence the page links rather than data it reports.
Questions about this entry
Short answers drawn from the linked project or page.
An open model from AutoTrust, fine-tuned from Qwen/Qwen3.5-9B and described on its card as a System 1 + System 2 model built with the Blocks of Experts recipe: one set of weights, one vLLM engine, routed per request.
Fast calibrated System 1 decisions plus System 2 generation and reasoning from an untouched Qwen3.5-9B. Since 3 October 2026 it also takes images, producing a probability for every action in a single forward pass in about 0.2 s on one GPU.
The card says its System 1 decisions are indistinguishable from the closed TypeSafe Jev 1.13 by KL, and that it answers a single decision in about a third of the time the hosted API takes. Its successor, autotrust/JEV-27B, is described as closer to Jev.
Five safetensors shards totalling about 8.95 GB in BF16 and F32, alongside transformer, lora, safetensors, and pytorch library entries, plus the SargeDev/jev-distill-corpus-v3 dataset as the training source.
The card shows a robot-arm pick-and-place from camera images, with System 1 answering questions such as whether the target is left or right of the gripper at every step. Run the vision path with bash vl/serve.sh.
A live recommendation and search demo at news.guru.so, new on 11 October 2026, where every ranking is a calibrated System 1 decision made in about half a second: headlines from six news feeds ranked by importance and search results re-ranked by relevance.
Licence apache-2.0, with metrics entries for kl, auroc, brier, and ece but no values attached in the captured page. Language is English, the organisation is AutoTrust, and the repository was uploaded from GitHub with 475 downloads.
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
X: Standard One SH (StandardThinking)
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
AutoTrust's bf16 System 1 decision model on Qwen3.5-35B-A3B with about 3B active parameters: Decision Index 0.3 public 59.73, ERQA 51.3% in one pass with zero robot training, 66.5 GiB of weights.
Jev Models & Reimplementations#community#open-models#local
Community
prithivMLmods' GGUF builds of autotrust/JEV-35B, the Qwen3.5-35B-A3B MoE decision model that returns calibrated choice, noul and score answers over text and images; files run from BF16 to Q3_K_L for llama.cpp.
Jev Models & Reimplementations#community#open-models#local
Community
Cloudflare's 9B multimodal decision model on Ollama: text, JSON, or images in one non-autoregressive forward pass, Apache-2.0, and the lowest measured latency of any decision model in Cloudflare's Decision Index runs at 38.8 ms median.
Jev Models & Reimplementations#community#open-models#local
Community
barakplasma's Replicate deployment of Jev-Omni: ask a typed yes/no or multiple-choice question about an image and get a calibrated probability per option from one forward pass, $0.00098 per run on L40S, Apache-2.0, and #1 on Image JevBench v0.1.
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
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
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.