CountingSheep/vev-4b

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 30, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

CountingSheep/vev-4b is a 4.5 billion parameter multimodal decision model developed by CountingSheep, built upon Qwen/Qwen3.5-4B. It is designed to answer yes/no, multiple-choice, or graded questions about text, JSON records, or images, providing probability scores for each answer. This model excels at visual question answering and structured decision-making, serving as a Jev-like model with added image understanding capabilities, making it suitable for tasks like UI state analysis or real-time game interaction.

Loading preview...

vev-4b: Multimodal Decision Model with Image Understanding

vev-4b is a 4.5 billion parameter multimodal decision model developed by CountingSheep, extending the Qwen/Qwen3.5-4B architecture. It specializes in answering structured questions (yes/no, multiple-choice, graded) about diverse inputs including text, JSON records, and images, providing probability distributions for all possible answers in a single forward pass. This model replicates the request format of TypeSafe's Jev API but adds crucial visual processing capabilities, allowing it to interpret screenshots and other visual data.

Key Capabilities

  • Multimodal Input Processing: Handles text, JSON, and images (screenshots) to answer questions.
  • Probabilistic Decision Making: Provides probability scores for each answer option, enabling nuanced decision support.
  • Jev-like API Compatibility: Uses the /v1/systemone request format, familiar to users of TypeSafe's Jev.
  • Visual Question Answering: Achieves high accuracy on tasks like identifying UI states in app screenshots (e.g., 0.928 for switch/checkbox status, 0.951 for text input fields).
  • Real-time Interaction: Demonstrated capability in scenarios like playing games by interpreting screenshots and making rapid decisions.

Good for

  • Automated UI Testing: Analyzing screenshots to determine application states or identify errors.
  • Content Moderation: Evaluating images against written safety policies.
  • Structured Data Extraction: Answering specific questions from text or JSON records with probabilistic confidence.
  • Interactive Agents: Developing agents that can perceive visual information and make structured decisions, such as in gaming or robotic control.
  • Research in Multimodal AI: Exploring decision-making processes in models that integrate visual and textual understanding.

Finetunes
2 models