OneRec-8B-pro

📖 Introduction

OpenOneRec is an open-source framework designed to bridge the gap between traditional recommendation systems and Large Language Models (LLMs). While Generative Recommendation has shown promise, existing models often struggle with isolated data silos and a lack of reasoning capabilities.

To address this, we introduce a unified framework that comprises:

• RecIF-Bench: The first holistic Recommendation Instruction-Following Benchmark, containing 100M interactions from 200k users across heterogeneous domains (Short Video, Ads, Product).
• OneRec-Foundation Models: A family of models (1.7B & 8B) built on the Qwen3 backbone. The series includes Standard versions trained on our open-source dataset and Pro versions enhanced with a hundred-billion-token industrial corpus from Kuaishou.
• Full-Stack Pipeline: We open-source our comprehensive training pipeline, including data processing, co-pretraining, and post-training, to ensure full reproducibility and facilitate scaling law research in recommendation.

🔥 News

• [2026.1.1] 📑 The technical report has been released.
• [2026.1.1] 🎉 OneRec-Foundation models (1.7B, 8B) are now available on Hugging Face!
• [2026.1.1] 🚀 RecIF-Bench dataset and evaluation scripts are open-sourced.
• [2026.1.5] 🔡 OneRec-Tokenizer is open-sourced to support SID generation for new domains.

📊 RecIF-Bench

We propose RecIF-Bench to rigorously assess the synergy between instruction following and domain-specific recommendation. It organizes 8 distinct tasks into a four-layer capability hierarchy:

• Layer 0: Semantic Alignment (Item Understanding)
• Layer 1: Fundamental Prediction (Short Video Rec, Ad Rec, Product Rec, Label Prediction)
• Layer 2: Instruction Following (Interactive Rec, Label-Conditional Rec)
• Layer 3: Reasoning (Recommendation Explanation)

The benchmark aggregates data from three domains: Short Video (Content), Ads (Commercial), and Product (E-commerce).

🤖 Model Zoo

The OpenOneRec-Foundation series is built upon the Qwen architecture, enhanced with Itemic Tokens for modality alignment and trained via a multi-stage protocol.

🏗️ Method & Architecture

OpenOneRec reframes recommendation as a general-purpose sequence modeling paradigm.

1. Items as Tokens

To bridge the modality gap, we treat items as a distinct modality using Itemic Tokens derived from hierarchical vector quantization. This allows the LLM to process interaction history as a cohesive context sequence.

2. Training Pipeline

Our framework utilizes the following recipe:

• Pre-Training: Integrates collaborative signals via Itemic-Text Alignment and Full-Parameter Co-Pretraining.
• Post-Training:
  • Stage 1: Multi-task Supervised Fine-tuning for basic instruction following.
  • Stage 2: On-policy Distillation to restore general reasoning performance.
  • Stage 3: Reinforcement Learning to enhance recommendation capabilities.

📈 Performance

Results on RecIF-Bench

OpenOneRec-Foundation achieves State-of-the-Art (SOTA) results across RecIF-Bench tasks, significantly outperforming baselines like LC-Rec and TIGER.

Cross-Domain Transferability

On the Amazon Benchmark (10 datasets), OpenOneRec demonstrates exceptional zero-shot/few-shot transfer capabilities, achieving an average 26.8% improvement in Recall@10 over the second-best method.

Metric: Recall@10. Ours refers to OneRec-Foundation with text-augmented itemic tokens strategy.

🚀 Quick Start

Code release and detailed usage instructions are coming soon.

Currently, you can load our models using transformers>=4.51.0:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "OpenOneRec/OneRec-8B"

# load the tokenizer and the model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# prepare the model input
# case - prompt with itemic tokens
prompt = "这是一个视频：<|sid_begin|><s_a_340><s_b_6566><s_c_5603><|sid_end|>，帮我总结一下这个视频讲述了什么内容"
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True # Switches between thinking and non-thinking modes. Default is True.
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
# Note: In our experience, default decoding settings may be unstable for small models.
# For 1.7B, we suggest: top_p=0.95, top_k=20, temperature=0.75 (during 0.6 to 0.8)
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

# parsing thinking content
try:
    # rindex finding 151668 (</think>)
    index = len(output_ids) - output_ids[::-1].index(151668)
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)

📜 Citation

If you find our work helpful, please cite our technical report:

@article{openonerec2025,
  title={An Open Foundation Model and Benchmark to Accelerate Generative Recommendation},
  author={OneRec Team},
  journal={arXiv preprint},
  year={2025}
}

🛡️ License

The code in this repository is licensed under the Apache 2.0 License. The model weights are subject to their specific license agreements.

🙏 Acknowledgements

OpenOneRec is built upon and inspired by the open-source ecosystem. We would like to thank:

• Qwen3: for providing the base architecture and model initialization that OpenOneRec builds upon.
• General-domain data sources: for the public corpora referenced in data/general_text used for mixed-domain training.
• VeRL & PyTorch distributed training: for the training infrastructure and scalable primitives (e.g., FSDP) used in post-training and large-scale runs.

We sincerely thank these projects for their outstanding work.