Visual decision model JEV-27B-VL ranks No. 1 on Hugging Face's global Models trending list; sister model GEV-26B-Decide ranks No. 3
JEV-27B-VL leads the Jev Decision Index vision board with a Full score of 69.8, 6.4 points above a 397-billion-parameter reference model
Open portfolio records 2.98 million downloads in 30 days as developers adopt integrated System 1 / System 2 decision models
SINGAPORE, Oct. 8, 2026 /PRNewswire/ -- AutoTrust AI Pte. Ltd. ("AutoTrust"), a Singapore-based applied AI research lab, today announced that JEV-27B-VL, its open-weight visual decision model, ranked No. 1 on Hugging Face's global Models trending list in rankings captured October 8, with its adaptive-reasoning model GEV-26B-Decide at No. 3.[1] JEV-27B-VL also leads the vision board of the Jev Decision Index, a public benchmark hosted on Hugging Face, with a Full score of 69.82, ahead of 67.66 for the next entrant.[2]
JEV-27B-VL at No. 1 and GEV-26B-Decide at No. 3 on Hugging Face’s Models trending list, in screenshots captured October 8, 2026. Rankings change over time.
Full results table: JEV-27B-VL leads on the Full score and on both public and private sets. Original screenshot; displayed values are rounded.
Jev Decision Index vision board, Full score (20 models plus 3 references, updated October 7, 2026). Original Hugging Face-hosted leaderboard screenshot, reproduced without edits.
Source: GEV-26B-Decide model card. Six public sets: AQuA-RAT, LogiQA, StrategyQA, MedMCQA, OpenBookQA and CommonsenseQA.
AutoTrust’s Hugging Face organization page. The 2.98 million figure is the rolling 30-day total across nine model repositories.
The results point to rising developer demand for what AutoTrust believes will become a distinct layer of the AI stack: decision models. Rather than generating paragraphs, a decision model reads a screen, image, document or record and returns a calibrated probability for each available action. AutoTrust's models pair a fast "System 1," which decides in a single forward pass, with a deliberate "System 2," which reasons step by step only when a call is uncertain. It is the fast-and-slow framework popularized by Nobel laureate Daniel Kahneman, delivered in a single open-model deployment.
"Language models taught AI to write. The bigger commercial prize is teaching AI to decide," said Josh Liu, Chairman and Co-Founder of AutoTrust AI. "Search, recommendation, advertising and fraud detection, the engines behind the internet's most profitable businesses, are decision systems at their core. Today they are stitched together from many narrow models that are costly to maintain and slow to adapt. A model that can see, score every option with calibrated confidence in milliseconds, and save expensive reasoning for the hard cases can consolidate much of that stack and make today's LLM deployments far more efficient. Reaching No. 1 on Hugging Face tells us developers see the same opportunity."
Key Highlights
Global developer traction. No. 1 (JEV-27B-VL) and No. 3 (GEV-26B-Decide) on Hugging Face's global Models trending list, and 2.98 million downloads across nine AutoTrust repositories in the rolling 30-day count, including 1.53 million for JEV-27B-VL.[1,6]
Benchmark leadership. JEV-27B-VL leads the Jev Decision Index vision board (20 models plus 3 references) on the Full score and on both the public (72.8) and private held-out (66.9) splits, and outscores the stock Qwen3.5-397B reference (63.4) by 6.4 points.[2]
Efficiency by design. GEV-26B-Decide's adaptive mode retains 99.5% of always-on reasoning accuracy (83.4% vs. 83.8%) while invoking System 2 on only 47.8% of 1,754 benchmark questions. System 1 decisions take about 45 milliseconds on a single NVIDIA B200 GPU.[4]
Evidence on real tasks. 95% completion on 60 multi-step browser tasks when element text is available; 15 of 20 simulated robot-arm pick-and-place scenes; and cold-start video recommendation that matches a collaborative-filtering baseline (AUC 0.727 vs. 0.728) without any interaction logs.[3]
Capital-efficient recipe. AutoTrust's Blocks of Experts recipe keeps a pretrained model frozen and adds a small, detachable decision block. JEV-9B's System 1 block has 40.2 million trained parameters, about 0.5% of its backbone, trained in roughly three B200 GPU-hours.[5]
Open and deployable. Open weights, with AutoTrust's components under Apache-2.0, served on vLLM through a single /v1/decide endpoint, with quantized variants available.[3,4,6]
The Investment Case for a Decision Layer
The first wave of generative AI was optimized for fluency. AutoTrust's thesis is that the next wave will be optimized for decisions, and that the economics favor models that integrate fast and slow thinking. A calibrated decision layer that sits beneath search, recommendation, advertising, fraud prevention, red teaming and agents, and in front of general-purpose LLMs, where it can route, verify and gate the expensive calls.
Where decision models fit
Workload
Typical approach today
Decision-model approach
AutoTrust evidence to date
Search and intent routing
Separate classifiers and rerankers per intent
One model scores up to 256 options per call
150-option intent questions in 1.4 s median (JEV-27B-VL)
Recommendation and ads
Models trained on click logs; weak on new items
Score new items from images and text before click history exists
Short-video AUC 0.727 vs. 0.728 for collaborative filtering trained on 59,045 users' logs
Fraud and risk
Rules engines and gradient-boosted trees
Calibrated yes/no risk calls; uncertain cases escalate to reasoning
Not yet publicly benchmarked
Red teaming, trust and safety
Separate guard models and LLM judges
Judge in one forward pass; reason on borderline cases
95% of 60 multi-step browser tasks (with element text)
Robotics
Hand-tuned policies or large end-to-end models
Image-to-action choice in about 240 ms
15 of 20 simulated pick-and-place scenes
AutoTrust evidence is self-reported on the JEV-27B-VL model card.[3] Typical approaches are generalized descriptions.
JEV-27B-VL: The Decision Engine Inside the Agent Loop
JEV-27B-VL is neither a chatbot nor a fully autonomous agent. It is the decision engine inside an agent loop. A developer sends a screenshot, camera frame or document with a question and a defined set of options. The model returns a calibrated probability for every option, such as "click this element," "route this case," "choose this answer" or "ask a human," instead of a paragraph that downstream software must interpret.[3]
System 1: yes/no checks, choices among 2 to 256 options, and 0-to-5 ratings in a single forward pass, with prompts of up to 256K tokens.
System 2: the unmodified base model reasons step by step, with image input, when a call warrants deeper thought.
One interface: a standard vLLM deployment exposes a /v1/decide endpoint, so teams can use one decision API instead of building a separate classifier for every workflow.
The most striking result is how JEV-27B-VL learned to see. Its System 1 decision adapter and scoring head were trained only on text, and its vision path is unchanged from its Qwen base model, so every image decision is zero-shot. Even so, it leads the vision board on both public and private held-out data, and it outscores a 397-billion-parameter stock vision-language reference by 6.4 points.[2,3] On a single NVIDIA B200, it processed 4,000 six-image decisions in 316 seconds, or about 13 per second.3 JEV-27B-VL is not a replacement for every large multimodal model. It is a control layer for systems that must make many small decisions quickly.
GEV-26B-Decide: Adaptive Thinking, Measured
GEV-26B-Decide makes the speed-versus-deliberation trade-off explicit. Built on Google's Gemma 4 26B-A4B mixture-of-experts model, with about 4 billion parameters active per token, it accepts text and images. System 1 answers first, in about 45 milliseconds on a B200 GPU, and sustains 257 decisions per second with 64 concurrent clients. When the leading option's probability is below 0.8, System 2 reasons and the two results are blended.[4]
Across 1,754 questions on six public benchmarks, adaptive mode scored 83.4%, compared with 73.3% for System 1 alone and 83.8% when reasoning on every question, while invoking System 2 on only 47.8% of them. Expected calibration error was 0.035. On the hardest sets, adaptive thinking lifted GPQA Diamond from 42.9 to 78.6 and MMLU-Pro from 65.0 to 84.6 relative to System 1 alone.[4]
A Fast-Growing Open Portfolio
Nine repositories on AutoTrust's Hugging Face organization page recorded 2.98 million downloads in the rolling 30-day count captured for this release.[6] Downloads are not revenue or customers. In open-source infrastructure, they are a leading indicator that engineers are testing, cloning and integrating the models. The highest-download models form a clear product ladder:
Model
Role
30-day downloads
JEV-27B-VL
Visual decisions for agents that inspect screens, documents, camera frames and image-heavy feeds
1.53 million
GEV-26B-Decide
Adaptive thinking: fast System 1 calls, with low-confidence cases escalated to System 2; text and image input
~896,000 (plus ~15,000 for the NVFP4 variant)
JEV-9B
Smallest model in the family, with the same decision API; browser and robot-arm simulation demos
~320,000
JEV-27B
Text-first 27B model whose System 1 adapter and decision head JEV-27B-VL builds on
~171,000
The road map has been shaped in public. After developers asked for computer use, 3D simulation and practical local deployment, AutoTrust added visual demos for shopping, settings, mail and robot-arm tasks, released quantized variants, and continued to publish serving code and evaluation reports.
Availability
JEV-27B-VL, GEV-26B-Decide, JEV-9B and JEV-27B are available now at huggingface.co/autotrust. AutoTrust's adapters, decision heads and calibration files are released under the Apache-2.0 license; base models remain under their original licenses. Enterprises interested in customized or sovereign decision models can contact AutoTrust at [email protected].
About AutoTrust AI
AutoTrust AI Pte. Ltd. is an applied AI research lab headquartered in Singapore, with an office in Silicon Valley. It builds the Guru family of foundation models; the JEV and GEV families of open decision models; and ScienceGuru, an agentic platform for scientific research and model development. AutoTrust also trains customized sovereign models for enterprises. Learn more at autotrust.ai and scienceguru.ai.
Media Contact
AutoTrust AI | autotrust.ai x.com/AutoTrustAI | huggingface.co/autotrust
Notes on Data
Hugging Face trending rankings and download counts are time-sensitive and change continuously; figures cited reflect screenshots captured October 8, 2026. Trending rank measures recent community attention, not model performance. Unless otherwise noted, AutoTrust model results are self-reported on AutoTrust's Hugging Face model cards. Results attributed to other companies are as published by those companies and have not been independently verified by AutoTrust. All trademarks are the property of their respective owners; references to third parties do not imply endorsement or affiliation.
Forward-Looking Statements
This press release contains forward-looking statements, including statements about AutoTrust's strategy, the development of decision models as a category, market opportunities, and the expected performance, adoption and applications of AutoTrust's models. Words such as "believe," "expect," "anticipate," "intend," "may," "will," "could" and similar expressions identify forward-looking statements. These statements reflect current expectations and are subject to risks and uncertainties, including rapid technological change, competition, the pace of developer and enterprise adoption, and the availability and cost of compute, any of which could cause actual results to differ materially. AutoTrust undertakes no obligation to update any forward-looking statement. This press release is for informational purposes only and does not constitute an offer to sell, or a solicitation of an offer to buy, any securities.
Share this article