Robot-arm simulation + computer use: JEV-9B now sees. Now multimodal: image-based System 1 decisions join its calibrated text decisions and System 2 reasoning. New community-requested visual demos: shopping, settings, mail, and a MuJoCo robot-arm simulation. AutoTrust leads open decision models. Model ↓ huggingface.co/autotrust/JEV… For AutoTrust AI’s full research-agent workspace, download ScienceGuru for Windows or macOS: scienceguru.ai/

Most decisions don't need a reasoning model. The hard ones do. GEV-26B-Decide knows the difference. It answers in ~45 ms, and stops to think only when it's unsure. What a moment of thought buys: 💣 Minesweeper: 21% → 86% 🔴 Connect Four: 54% → 99.5% 🟩 Wordle: 52% → 100% 🔢 Sudoku: 76% → 99.5% JEV Decision Index: 53.3 for our JEV-27B-VL → 62.48 for GEV-26B-Decide, even higher than the current leader, Cloudflare's Clef (61.2). All self-reported. Open weights 👇 huggingface.co/autotrust/GEV…

AutoTrust AI Lab, in motion. JEV-9B, JEV-27B, JEV-27B-VL & JEV-Gemma4-26B-A4B pair fast probabilistic decisions with reasoning. GEV-26B-Decide adds adaptive thinking. SLIM-Q GLM-5.3/Flash GGUF builds target DGX Spark. Watch the models play ↓ huggingface.co/autotrust

Thanks @Gorden_Sun, @potechi_takusan, @Read_Acted, and @a_captaincook — the questions have been great. A few clear answers: • Computer use: yes. The path from screen state → calibrated next action → click / type / scroll loop is a core direction for JEV-27B-VL. • 3D + robotics: also yes. The current release takes 2D image state; 3D scene understanding and embodied / robotics workflows are directions we are actively building toward, not capabilities we claim are fully solved today. • Local deployment: agreed — an 80GB GPU is too much for many builders. The current BF16 release targets that class of hardware, but quantized and GGUF variants are coming soon. • For agent engineers: JEV-27B-VL is open-weight under Apache-2.0, with vLLM and SGLang deployment paths. It is designed for calibrated visual decisions, not just image description. Mario and Tetris make the loop easy to see: see → decide → act. The destination is broader: computer use, visual agents, and eventually embodied systems. JEV-27B-VL: huggingface.co/autotrust/JEV… For AutoTrust AI’s full research-agent workspace, download ScienceGuru for Windows or macOS: scienceguru.ai/

JEV-27B-VL：开源多模态版本Jev模型 像人类一样有两种思考模式： 第一种是“直觉反应”（快思考）。面对选择题、判断对错或者打分这类任务，它看一眼文字或图片，不用逐字生成长篇大论，一次计算就能直接给出每个选项的概率。比如判断一张照片是不是食物，或者在几十个分类里挑出最合适的一项，反应速度只需要几毫秒到几百毫秒。 第二种是“深度分析”（慢思考）。当遇到复杂的图表或需要深入推导的逻辑题时，它会切换到传统的分析模式，一步一步梳理细节并输出完整的文本解释。 支持图片输入，能用于视频推荐、玩游戏等场景。 模型：huggingface.co/autotrust/JEV…

Thanks, @mr_r0b0t — really glad you enjoyed it! 🤩 The Mario demo is exactly the kind of loop we wanted to make tangible: visual state in, calibrated action probabilities out, then the agent acts. JEV-27B-VL is open-weight under Apache-2.0, so we’re excited to see what people build with it.

A week of JEV at AutoTrust AI: from fast text decisions to open multimodal action loops. JEV is our System 1 family for agents. Instead of generating a paragraph before every small action, JEV returns calibrated probabilities for yes/no, 2–16-way choice, and 0–5 scoring — so an agent can act immediately, or escalate hard cases to System 2 reasoning. The foundation is our Blocks of Experts (BoE) recipe: keep a capable base model intact, add a small decision expert, and route each request to the right path. Fast decisions and deeper reasoning, served from one engine. What we have released: • JEV-9B — our first integrated open System 1 + System 2 model. It reaches ≈0.019 KL to TypeSafe Jev 1.13’s decision distributions, while making a decision in about 90 ms median on one B200. Its System 1 block trains only 40.2M parameters — 0.5% of the backbone. • JEV-27B — our stronger 27B model. It reaches ≈0.017 KL, achieves 96% of Jev’s accuracy on an independent 16-option benchmark, and scores 84.07% across six public decision benchmarks in our updated evaluation, ahead of Jev 1.13’s 83.85% in 4 of 6 groups. Its System 2 path remains unchanged at 78.0% HumanEval, while System 1 runs at 137 ms median per decision on one B200. • JEV-Gemma4-26B-A4B — a BoE System 1 + System 2 bundle on Gemma 4 MoE. It scored 58.05 on Decision Index 0.2.1, completing all 150,317 scored requests with 0 errors and 0 unsupported. • JEV-27B-VL — what we believe is the world’s first open-weight, near-SOTA multimodal decision model. It extends calibrated decisions to images and visual state: Mario, Tetris, and Rubik’s Cube are not chat demos; they are see → decide → act loops. Across the JEV family, the goal is the same: give agents a fast, calibrated System 1 for high-frequency judgments — without giving up the System 2 reasoning needed for hard work. Explore the JEV family, models, demos, and evaluation results: huggingface.co/autotrust Then use these capabilities inside a full research workspace. Download ScienceGuru for Windows or macOS: scienceguru.ai/

Decision models are becoming core infrastructure for agents. JEV-27B is our open-weight System 1: calibrated yes/no, 2–16-way choice, and 0–5 scoring in one forward pass — 137 ms median per decision on a single B200, measured server-side without network latency. Today, we add vision with JEV-27B-VL — an open-weight multimodal decision model built with AutoTrust AI’s Blocks of Experts (BoE) recipe. It makes calibrated yes/no, 2–16-way choice, and 0–5 scoring decisions directly over text and images, while keeping System 2 reasoning in the same engine. On MicroLens-100k, its zero-shot, cover-only recommendation reaches 0.727 AUC: on par with collaborative filtering trained on 59,045 users’ watch histories, with a higher HR@5. Updated model card: huggingface.co/autotrust/JEV…

AI agents live or die by the small decisions: select, route, verify, act. We’ve released the complete Decision Index 0.2.1 outputs for JEV-27B on @huggingface. • 53.30 Decision Index score • 64.32 raw score • 150,317 / 150,317 scoreable requests completed • 0 errors, 0 unsupported requests This is more than a headline number. The open result set includes compact answers, probabilities, timings, statuses, environment records, engine code, hashes, and training-overlap checks—so the run can be inspected and re-scored with the Decision Index kit. JEV-9B results are included too: 43.14 → 53.30 on the same Decision Index edition. Explore the open results: huggingface.co/datasets/auto…