CountingSheep/vev-9b

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 30, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

CountingSheep/vev-9b is a 9-billion parameter multimodal decision model developed by CountingSheep, built upon Qwen/Qwen3.5-9B. It is designed to answer yes/no, multiple-choice, or graded questions about text, JSON records, or images, providing probabilities for allowed answers in a single forward pass. This model excels at visual question answering and decision-making tasks, supporting a 32768-token context length and offering Jev-like functionality with added image understanding capabilities.

Loading preview...

Overview

CountingSheep/vev-9b is a 9-billion parameter multimodal decision model, extending the capabilities of Jev-like models by incorporating image understanding. Built on Qwen/Qwen3.5-9B, it processes text, JSON records, and images to answer yes/no, multiple-choice, or graded questions, outputting probabilities for each possible answer. This model is particularly suited for applications requiring rapid, probabilistic decision-making based on diverse input types, including visual information.

Key Capabilities

  • Multimodal Input: Processes text, JSON, and images (screenshots, photos) for decision-making.
  • Probabilistic Answers: Provides a probability for every allowed answer in a single forward pass, supporting yes/no (Noul), multiple-choice (Choice), and graded questions.
  • Jev API Compatibility: Serves the same request format as TypeSafe's Jev (/v1/systemone), allowing for seamless integration into existing workflows.
  • Visual Question Answering: Demonstrates strong performance on tasks like UI state analysis in app screenshots (0.923-0.955 accuracy), general photo questions (MMBench-EN 0.902), and image-to-policy compliance checks (LlavaGuard 0.724).
  • Text-based Decision Making: Achieves competitive accuracy on Jev-style decisions on text (JevBench public subset 0.823) and Chinese text (judgekit 0.938).

Good For

  • Automated UI Testing & Monitoring: Identifying elements, states, or errors in app screenshots.
  • Content Moderation & Policy Enforcement: Evaluating images against safety policies or prompts.
  • Customer Support Automation: Analyzing customer queries (text) and associated visual context (screenshots) to determine urgency or next steps.
  • Game AI & Automation: Real-time decision-making based on game screenshots, as demonstrated by vev-4b playing Doom.
  • Any application requiring rapid, probabilistic decision-making from diverse data sources, including visual data.

Limitations

  • Probabilities are not exact frequencies, with expected calibration error between 0.01–0.14.
  • Order of options can influence top answer in some cases (13% on JevBench).
  • Less effective than the base model for multi-step reasoning tasks where the base model can "think first".

Finetunes
2 models