juspay/Xor-26B-A4B
juspay/Xor-26B-A4B is a 26 billion parameter language model, post-trained by Juspay from Google's Gemma 4 26B-A4B-it base, with a 32768 token context length. It is specifically designed and optimized for typed decision tasks, handling text, image, and video inputs to answer 'noul', 'choice', and 'score' questions. This model excels at structured decision-making rather than general chat or prose generation, making it suitable for applications requiring precise, categorized responses.
Loading preview...
Xor-26B-A4B v1.2 Overview
Xor-26B-A4B v1.2 is a 26 billion parameter model developed by Juspay, post-trained from Google's Gemma 4 26B-A4B-it base. This version is a weight blend of two post-trained checkpoints, specifically optimized for typed decision tasks. It processes text, image, and video inputs to answer structured questions like 'noul' (yes/no), 'choice' (multiple options), and 'score' (ordered levels).
Key Capabilities
- Specialized Decision-Making: Designed for precise, structured responses to specific question types, not general chat.
- Multimodal Input: Accepts text, up to 8 base64 image data URLs, or one base64 video data URL per request.
- Optimized for Structured Output: The wrapper scores candidate label-token probabilities, supporting 2-255 options for 'choice' and 2-10 ordered levels for 'score'.
- High Context Length: Supports a context limit of 131,072 tokens, accommodating substantial input.
- Performance: Achieved a Public Decision Index score of 59.46 in self-run evaluations, demonstrating efficiency in decision tasks.
Good For
- Automated Decision Systems: Ideal for applications requiring clear, categorized decisions based on diverse inputs.
- Structured Data Extraction: Useful for extracting specific types of information or making classifications from text, images, or video.
- High-Throughput Inference: Designed for efficient serving, with reported mean HTTP request latency of 559.5 ms under load on specific hardware.
- Applications Requiring Precise Answers: Suited for use cases where general prose generation is not needed, but accurate, type-constrained responses are critical.