AIAny
AI Model2026
Icon for item

autotrust/JEV-27B-VL

Delivers fast, calibrated 'System 1' decisions over text and images (yes/no, 2–256 choices, 0–5 scores) in one forward pass, plus full 'System 2' Qwen3.8-27B multimodal reasoning. Highlights: zero-shot vision judging, 256K-token context, vLLM decision API, Apache-2.0.

Introduction

Most real-world agents need many quick, well-calibrated decisions rather than long-form generation. JEV-27B-VL separates a fast System 1 decision head (LoRA adapter) that returns calibrated probabilities in one forward pass from an untouched Qwen3.8-27B System 2 for deliberate multimodal reasoning — and extends both modes to image inputs zero-shot.

Key Capabilities
  • Fast, calibrated decisions: yes/no (noul), multi-way choice (2–256 options), and 0–5 scoring in a single forward pass, with very low calibration error and high fidelity to TypeSafe Jev distributions.
  • Multimodal support: decision head applied zero-shot to images (System 1) and full image-aware generation/reasoning via the Qwen3.8-27B backbone (System 2).
  • Production-friendly API and scale: serves via vLLM with a dedicated POST /v1/decide endpoint, supports up to 256K-token prompts, LoRA adapters for System 1, and Apache-2.0 weights.
  • Proven applied strengths: top multimodal judge on VL-RewardBench (78.3% accuracy), zero-shot short-video recommendation from covers (AUC 0.727, matching collaborative filtering), strong agent-judging and search re-ranking results.
Who it's for and trade-offs

Great fit if you need high-throughput, calibrated automated judgments over text and images (content moderation, recommendation ranking, UI action selection, agent evaluation) and want to combine these with a full reasoning LLM in the same deployment. Look elsewhere if you require fully supervised, image-trained decision heads (image decisions here are zero-shot and their calibration is less systematically measured) or if you cannot allocate GPU memory for large-context multimodal runs (full 256K context raises memory requirements). The model trades occasional extra latency for very wide context and a single unified weights+adapter deployment.

Where it fits

JEV-27B-VL is positioned between specialized closed hosted decision services and large multimodal generative models: it delivers calibrated decision APIs suitable for real-time ranking/control loops while retaining the full Qwen3.8-27B reasoning ability for escalation to System 2. Use it when you need deterministic, interpretable probabilities at scale and an integrated path to deeper multimodal explanations.

Information

  • Websitehuggingface.co
  • Organizationsautotrust, Qwen
  • Published date2026/09/30

Categories

More Items

Hugging Face
AI Model2026

Generates 16×16 pixel-art sprites from text by representing the canvas as 256 palette-color pixels and filling them with masked diffusion in parallel. Non-commercial use only unless separately licensed.

Hugging Face
AI Model2026

Uses reinforcement learning alone to improve Qwen3.8-27B on multimodal reasoning and long-horizon agent tasks. Trained on about 1,200 curated tasks, it retains a 256K context and is released in BF16, FP8, and NVFP4.

Hugging Face
AI Model2026

Transcribes speech audio into text across 45 languages, with automatic language detection. Its 0.5B parameters make it a compact option for multilingual transcription using Transformers or vLLM.