Ouro-2.6B-Thinking

Model Description

**⚠️ IMPORTANT: This model is intended for research purposes only. It is provided as-is without warranties for production use. **

Ouro-2.6B-Thinking is a reasoning-specialized variant of the Ouro-2.6B base model, enhanced through supervised fine-tuning on high-quality reasoning data. Please use transformers==4.54.1for compatibility.

Key Features

• Advanced Reasoning: Specifically optimized for mathematical and scientific reasoning tasks
• Compact Size: Competitive with 4B models despite having only 2.6B parameters
• Cross-Step Consistency: Intermediate recurrent outputs can serve as reliable proxies for final answers
• Explicit Thinking Process: Trained to generate detailed reasoning steps

Configuration

Recurrent Steps and Adaptive Exit

The model's computational behavior can be configured through the config.json file:

{
  "total_ut_steps": 4,
  "early_exit_threshold": 1.0
}

• total_ut_steps: Controls the number of recurrent steps (default: 4). You can adjust this value to trade off between performance and computation time.
• early_exit_threshold: Controls the adaptive exit mechanism (default: 1.0). Lower values encourage earlier exit, while 1.0 means always use all steps.

Example: Modify recurrent steps

from transformers import AutoConfig, AutoModelForCausalLM

config = AutoConfig.from_pretrained("ByteDance/Ouro-2.6B-Thinking")
config.total_ut_steps = 3  # Use 3 recurrent steps instead of 4
model = AutoModelForCausalLM.from_pretrained(
    "ByteDance/Ouro-2.6B-Thinking",
    config=config,
    device_map="auto"
)

Note: vLLM does not currently support the adaptive exit feature due to its inference optimization characteristics. When using vLLM, the model will always execute the full number of total_ut_steps.

Model Architecture

Based on Ouro-2.6B with additional reasoning fine-tuning:

Configuration	Value
Parameters	2.6B
Layers	24
Recurrent Steps	4
Hidden Size	2048
Attention Heads	Multi-Head Attention (MHA)
FFN Activation	SwiGLU
Position Embedding	RoPE
Vocabulary Size	49,152
Context Length	32K (SFT)
Normalization	Sandwich RMSNorm

Training Details

Pre-training

• Training Tokens: 7.7T tokens across 4 stages
• Base Architecture: Ouro-2.6B

Supervised Fine-Tuning

• Data Size: ~8.3M examples
• Data Composition:
  • Mathematics: 3.5M examples (OpenThoughts3, AceReason-1.1-SFT)
  • Code: 3.2M examples (AceReason, OpenCodeReasoning, Llama-Nemotron, OpenThoughts3)
  • Science: 808K examples (OpenThoughts3, Llama-Nemotron)
  • Chat: 767K examples (DeepWriting-20K)
• Training: 2 epochs, max sequence length 32K
• Optimizer: Adam (lr=2×10⁻⁵, β=(0.9, 0.95))
• Scheduler: Cosine decay

Quick Start

⚠️ IMPORTANT: Please use transformers<4.56.0 to avoid compatibility issues. We recommend transformers==4.54.1 or earlier versions.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Bytedance/Ouro-2.6B-Thinking"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype="auto"
)

# Generate with reasoning
messages = [
    {"role": "user", "content": "Solve: If 2x + 3 = 11, what is x?"}
]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(inputs, max_new_tokens=512, temperature=1.0, top_p=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citation

@article{ouro2025,
  title={Scaling Latent Reasoning via Looped Language Models},
  author={Zhu, Rui-Jie and Wang, Zixuan and Hua, Kai and Zhang, Tianyu and Li, Ziniu and Que, Haoran and Wei, Boyi and Yin, Fan and Wen, Zixin and Xing, He and others},
  journal={arXiv preprint},
  year={2025}
}

License

This model is licensed under Apache-2.0. See the LICENSE file for details.

Project Links

• Paper: Scaling Latent Reasoning via Looped Language Models
• Project Page: https://ouro-llm.github.io

---