KartiOS/Karti-Small-VL-4B
KartiOS/Karti-Small-VL-4B is a 4.5 billion parameter vision-language model built on Qwen/Qwen3.5-4B, specifically designed for local agents that interact with visual information. This model excels at accurately reading panels, checking camera frames, and acting on screens by significantly reducing confidently invented identifiers. It is optimized to decline when it cannot reliably read something, making it highly reliable for tool-calling applications where false positives are critical.
Loading preview...
Karti-Small-VL-4B: A Vision Model for Reliable Local Agents
Karti-Small-VL-4B is a 4.5 billion parameter vision-language model, fine-tuned from Qwen/Qwen3.5-4B, with a 32,768 token context length. Its core innovation lies in its ability to reliably decline when it cannot accurately read visual information, drastically reducing the rate of "confidently invented identifiers" that plague many vision models. This makes it particularly suitable for applications where incorrect outputs are costly, such as controlling local agents that interact with user interfaces or physical environments.
Key Capabilities & Performance
- Reduced Invented Identifiers: Achieves a 16x lower invented-identifier rate (0.023 vs 0.378 for base model), ensuring higher reliability.
- Enhanced Legibility: Significantly improves panel reading (0.967 vs 0.595) and GUI grounding on ScreenSpot-v2 (0.907 vs 0.718), especially for small targets.
- Tool Integration: Features native tool support with a
qwen3_xmlparser, designed for agentic workflows. - Efficient Deployment: Available in BF16 precision (9.3 GB) and an NVFP4 build for 2.5x faster throughput on Blackwell GPUs (3.7 GB).
Training & Focus
The model was trained using BF16 LoRA, with the vision tower frozen, focusing on a curated corpus that included a "legibility ladder" and datasets like VizWiz and GUI-Odyssey. A key training insight was balancing the proportion of "I can't read this" answers to teach the model judgment rather than just the phrase. While it excels at rendered panels and screens, its performance on general photographic vision remains similar to the base model, with the improvement being its ability to avoid over-refusing them.
Ideal Use Cases
- Local Agents: Perfect for agents that need to interpret visual data from screens, cameras, or panels and then call external tools.
- UI Automation: Reliably extracting information from graphical user interfaces without hallucinating non-existent elements.
- Robotics/Embedded Vision: Scenarios where accurate, non-hallucinated visual parsing is critical for decision-making.