ByteDance/Ouro-1.4B-Thinking

TEXT GENERATIONConcurrency Cost:1Model Size:1.4BQuant:BF16Ctx Length:32kPublished:Oct 28, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

ByteDance/Ouro-1.4B-Thinking is a 1.4 billion parameter causal language model developed by ByteDance, specifically fine-tuned for advanced reasoning tasks. This model excels in mathematical and scientific problem-solving, generating explicit reasoning steps, and demonstrates performance competitive with larger 4B models. It features a 32K context length and a recurrent architecture for cross-step consistency, making it suitable for complex analytical applications.

Loading preview...

Ouro-1.4B-Thinking: A Reasoning-Specialized LLM

Ouro-1.4B-Thinking is a 1.4 billion parameter language model from ByteDance, built upon the Ouro-1.4B base model and enhanced through supervised fine-tuning on high-quality reasoning data. This model is designed for research purposes and focuses on advanced analytical capabilities.

Key Capabilities

  • Advanced Reasoning: Optimized for complex mathematical and scientific reasoning tasks, generating detailed, explicit reasoning steps.
  • Compact Efficiency: Achieves performance comparable to models with 4 billion parameters despite its smaller 1.4 billion parameter count.
  • Cross-Step Consistency: Utilizes a recurrent architecture (default 4 steps) where intermediate outputs are reliable proxies for final answers.
  • Configurable Recurrence: Allows adjustment of total_ut_steps and early_exit_threshold via config.json to balance performance and computation.

Training Details

The model underwent pre-training with 7.7T tokens and subsequent supervised fine-tuning on approximately 8.3 million examples. The fine-tuning dataset composition includes 3.5M mathematics examples (OpenThoughts3, AceReason-1.1-SFT), 3.2M code examples, and 808K science examples, trained for 2 epochs with a max sequence length of 32K.

Good For

  • Applications requiring strong mathematical and scientific problem-solving.
  • Scenarios where explicit, step-by-step reasoning is beneficial.
  • Environments needing a compact yet powerful model for reasoning tasks.