JEV-27B-VL
A 27B multimodal decision model that ranks and recommends with calibrated probabilities.
About
JEV-27B-VL is a vision-language decision model from AutoTrust AI Lab, a small applied lab of about ten researchers working on agentic models, released in late September 2026 under Apache 2.0. It is a LoRA/adapter fine-tune on Qwen3.8-27B, distilled on the SargeDev/jev-distill-corpus-v3 dataset. Instead of producing rankings or chat, it makes typed System-1/System-2 decisions with calibrated probabilities: the intended use is replacing a ranking team with one decision model for recommendation, search re-ranking, and embodied reasoning. AutoTrust's headline claim is 61.5% on ERQA embodied reasoning with thinking enabled, which they describe as the top open-model result with zero robot training.
The live proof is GuruSearch (news.guru.so), a demo launched October 11, 2026 that ranks headlines from six news feeds and re-ranks Google/Bing/Yahoo News results in about half a second per decision, with a short cited answer. The ERQA figure and the "one decision model instead of a ranking team" framing are provider claims, documented in the Hugging Face blog post; the checkpoint, adapter lineage on Qwen3.8-27B, Apache 2.0 license, and the live demo are independently confirmable. As a LoRA on an Apache-licensed base, it inherits clean licensing for commercial use.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: Verified via Hugging Face model card (autotrust/JEV-27B-VL): LoRA/adapter on Qwen/Qwen3.8-27B, Apache 2.0, distilled on SargeDev/jev-distill-corpus-v3, typed System-1/System-2 decisions with calibrated probabilities.; Weights confirmed downloadable on HF (ungated); repo created 2026-09-30; 5.2k likes (high for a 2-week-old model); HF blog post documents the approach; provider is AutoTrust AI Lab (~10 researchers, per HF org page).; ERQA 61.5% 'top open model' claim is AutoTrust's own with no independent rerun. Live GuruSearch demo (news.guru.so, launched 2026-10-11) is independently reachable in principle but was not exercised end-to-end.. License: Apache-2.0 (model card frontmatter); base Qwen3.8-27B also Apache-2.0.