KAT-V1-40B

<div align="center"> <img src="https://raw.githubusercontent.com/Anditty/OASIS/refs/heads/main/Group.svg" width="60%" alt="Kwaipilot" /> </div> <hr> <div align="center" style="line-height: 1;"> <a href="https://huggingface.co/Kwaipilot/KAT-V1-40B" target="_blank"> <img alt="Hugging Face" src="https://img.shields.io/badge/HuggingFace-fcd022?style=for-the-badge&logo=huggingface&logoColor=000&labelColor"/> </a> <a href="https://arxiv.org/pdf/2507.08297" target="_blank"> <img alt="arXiv" src="https://img.shields.io/badge/arXiv-2507.08297-b31b1b.svg?style=for-the-badge"/> </a> </div>

News

• We released the technical report of the KAT-V1 model, available at https://arxiv.org/pdf/2507.08297.
• Kwaipilot-AutoThink ranks first among all open-source models on LiveCodeBench Pro, a challenging benchmark explicitly designed to prevent data leakage, and even surpasses strong proprietary systems such as Seed and o3-mini.

---

Introduction

KAT (Kwaipilot-AutoThink) is an open-source large-language model that mitigates over-thinking by learning when to produce explicit chain-of-thought and when to answer directly.

Its development follows a concise two-stage training pipeline:

<table> <thead> <tr> <th style="text-align:left; width:18%;">Stage</th> <th style="text-align:left;">Core Idea</th> <th style="text-align:left;">Key Techniques</th> <th style="text-align:left;">Outcome</th> </tr> </thead> <tbody> <tr> <td><strong>1. Pre-training</strong></td> <td>Inject knowledge while separating “reasoning” from “direct answering”.</td> <td> <em>Dual-regime data</em><br> • <strong>Think-off</strong> queries labeled via a custom tagging system.<br> • <strong>Think-on</strong> queries generated by a multi-agent solver.<br><br> <em>Knowledge Distillation&nbsp;+&nbsp;Multi-Token Prediction</em> for fine-grained utility. </td> <td>Base model attains strong factual and reasoning skills without full-scale pre-training costs.</td> </tr> <tr> <td><strong>2. Post-training</strong></td> <td>Make reasoning optional and efficient.</td> <td> <em>Cold-start AutoThink</em> — majority vote sets the initial thinking mode.<br> <em>Step-SRPO</em> — intermediate supervision rewards correct <strong>mode selection</strong> and <strong>answer accuracy</strong> under that mode. </td> <td>Model triggers CoT only when beneficial, reducing token use and speeding inference.</td> </tr> </tbody> </table>

---

Data Format

KAT produces responses in a structured template that makes the reasoning path explicit and machine-parsable.
Two modes are supported:

Special Tokens

---

🔧 Quick Start

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "Kwaipilot/KAT-V1-40B"

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

# prepare the model input
prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# conduct text completion
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=65536,
    temperature=0.6,
    top_p=0.95,
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 
content = tokenizer.decode(output_ids, skip_special_tokens=True).strip("\n")
print("prompt:\n", prompt)
print("content:\n", content)
"""
prompt:
Give me a short introduction to large language model.
content:
<judge>
The user's request is to provide a concise factual introduction to large language models, which involves retrieving and summarizing basic information. This task is straightforward as it only requires recalling and presenting well-known details without deeper analysis. No complex reasoning is needed here—just a simple explanation will suffice.
</judge>

<think_off>
<answer>
A **Large Language Model (LLM)** is an advanced AI system trained on vast amounts of text data to understand, generate, and process human-like language. Here’s a concise introduction:

### Key Points:
1. **Training**: Trained on diverse text sources (books, websites, etc.) using deep learning.
2. **Capabilities**: 
   - Answer questions, generate text, summarize content, translate languages.
   - Understand context, sentiment, and nuances in language.
3. **Architecture**: Often based on **transformer models** (e.g., BERT, GPT, LLaMA).
4. **Scale**: Billions of parameters, requiring massive computational resources.
5. **Applications**: Chatbots, content creation, coding assistance, research, and more.

### Examples:
- **OpenAI’s GPT-4**: Powers ChatGPT.
- **Google’s Gemini**: Used in Bard.
- **Meta’s LLaMA**: Open-source alternative.

### Challenges:
- **Bias**: Can reflect biases in training data.
- **Accuracy**: May hallucinate "facts" not grounded in reality.
- **Ethics**: Raises concerns about misinformation and job displacement.

LLMs represent a leap forward in natural language processing, enabling machines to interact with humans in increasingly sophisticated ways. 🌐🤖
</answer>
"""

---

Future Releases

Looking ahead, we will publish a companion paper that fully documents the AutoThink training framework, covering:

• Cold-start initialization procedures
• Reinforcement-learning (Step-SRPO) strategies
• Data curation and reward design details

At the same time, we will open-source:

• Training resources – the curated dual-regime datasets and RL codebase
• Model suite – checkpoints at 1.5B, 7B, and 13B parameters, all trained with AutoThink gating

Citation

@techreport{Zhan2025KATV1,
  title={KAT-V1: Kwai-AutoThink Technical Report},
  author={Zizheng, Zhan and Ken, Deng and Huaixi, Tang and Wen, Xiang and Kun, Wu and others},
  year={2025},
  institution={arXiv preprint arXiv:2507.08297},
  number={arXiv:2507.08297},
  url={https://arxiv.org/abs/2507.08297}
}