juspay/xor

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 21, 2026License:apache-2.0Architecture:Transformer0.1K Open Weights Featherless Exclusive Cold

Juspay's Xor 1.2 is a 35.1 billion parameter Mixture-of-Experts causal language model, post-trained from Qwen/Qwen3.6-35B-A3B, specifically optimized for typed decision tasks. It excels at classifying information from text and media inputs, providing binary probabilities, categorical decisions, and ordinal scores. This model is designed for reproducible results in structured decision-making workflows, particularly with its TypeSafe-compatible API and validated SGLang runtime.

Loading preview...

Xor 1.2: A Specialized Model for Typed Decision Tasks

Xor 1.2 is a 35.1 billion parameter Mixture-of-Experts (MoE) causal language model developed by Juspay, building upon the Qwen/Qwen3.6-35B-A3B base. This version is post-trained with LoRA (rank 16) specifically for typed decision tasks, offering a TypeSafe-compatible /v1/systemone API.

Key Capabilities

  • Typed Decision Outputs: Provides structured outputs for decision-making, including binary probabilities (noul), categorical decisions with full probability distributions (choice), and expected ordinal scores with distributions (score).
  • Multimodal Input Processing: Capable of processing text alongside up to eight image data URLs or one video data URL, allowing typed questions to classify information from various media types.
  • Reproducible Inference: Features a validated serving layer that performs deterministic single-token candidate readout, option-order evaluation, probability calibration, and schema conversion, ensuring consistent results.
  • Performance: Achieved a macro accuracy of 0.9070 across all public tiers of the JEVBench benchmark in self-run evaluations.

Good For

  • Structured Decision-Making: Ideal for applications requiring precise, typed outputs for decisions based on complex inputs.
  • Multimodal Classification: Suitable for use cases where classification or decision logic needs to incorporate both textual and visual information.
  • Reproducible AI Workflows: Recommended for environments where consistent and verifiable model behavior is critical, thanks to its pinned revisions and validated runtime.

Finetunes
6 models