InfiFusion-14B

<h1 align="center"> InfiFusion: A Unified Framework for Enhanced Cross-Model Reasoning via LLM Fusion. </h1> <h4 align="center">

</h4>

InfiFusion is the first fusion framework for large language models that fuse up to 4 models with 14B~24B parameters. We introduce a unified framework which can fuse many heterogeneous models in one distillation stage. InfiFusion outperforms the state-of-the-art models, such as Qwen-2.5-14B-Instruct and Phi-4, across 11 widely applied benchmarks covering reasoning, coding, mathematics, and instruction-following tasks. Notably, InfiFusion achieves this superior performance while significantly reduces computational costs, completing full training with only 160 H800 GPU hours compared to the millions typically required for traditional LLM training.

🎉 News

🎉 We've retrain infiFusion using new data (refer to Training Data). The ckpt model, InfiFusion-14B, has been released on Huggingface! ! !

🎨 Fusion Framework

<img src="https://cdn-uploads.huggingface.co/production/uploads/676279cf57e77fe6e0d3a09a/1hpClyRzIk9UnT8XIKU71.png" width="800">

📕 Model Summary

🩺 Intended Use

💼 Data Overview

📚 Training Data

We construct a novel multi-task training dataset comprising 130k curated examples across three major domains: general reasoning, mathematics, and code generation.

1. General Reasoning (52K samples) Samples are sourced from the Infinity-Instruct dataset, a high-quality instruction-following corpus created through expert filtering.

2. Mathematics (39K samples) Questions are drawn from the NuminaMath-1.5 dataset—an advanced benchmark for competition-level math spanning Algebra, Geometry, Combinatorics, Calculus, Inequalities, Logic & Puzzles, and Number Theory. Answers are distilled from the DeepSeek-R1-671B model by the AM team.

3. Code Generation (39K samples) We used KodCode-V1-SFT-R1, a dataset with 268K code samples. Each example was processed by our pivot model to generate five completions. These were sandbox-evaluated, and samples where at least one generation failed were flagged. From these, we filtered and distilled 39K high-quality examples.

Benchmark evaluation

To enhance the robustness of answer extraction under the regex-based evaluation framework ofOpenCompass and EvalPlus, we systematically refine the prompts used in several benchmark datasets. These tailored prompt formats are designed to facilitate precise output matching, mitigating ambiguities that often arise from model generations. The revised prompt templates corresponding to each dataset are presented in the following Table, which details how task instructions and answer formats are standardized to align with OpenCompass's automatic evaluation pipeline.

For datasets such as TheoremQA and HumanEval, we retain the original prompt configurations, adhering to their respective community-adopted evaluation protocols. This ensures consistency with prior works and preserves the validity of established benchmarks. For MBPP, we utilize EvalPlus for a more rigorous assessment of LLM-generated code, providing enhanced reliability in functional correctness evaluation, more specifically:

🚀 Usage

Input Formats

Our fusion process uses phi-4 as the pivot model, and thus InfiFusion shares the same prompt format and usage style as phi-4.

Given the nature of the training data, InfiFusion performs best when used with prompts in the following chat-style format:

<|im_start|>system<|im_sep|>
You are a medieval knight and must provide explanations to modern people.<|im_end|>
<|im_start|>user<|im_sep|>
How should I explain the Internet?<|im_end|>
<|im_start|>assistant<|im_sep|>

With transformers

import transformers

pipeline = transformers.pipeline(
    "text-generation",
    model="InfiFusion",  # replace with actual model path
    model_kwargs={"torch_dtype": "auto"},
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are a medieval knight and must provide explanations to modern people."},
    {"role": "user", "content": "How should I explain the Internet?"},
]

outputs = pipeline(messages, max_new_tokens=1024)
print(outputs[0]["generated_text"][-1])

📌 Note: Since InfiFusion uses phi-4 as its pivot model during fusion, it inherits many of its usage patterns and prompt compatibility features.

🎯 Model Quality

<img src="https://cdn-uploads.huggingface.co/production/uploads/676279cf57e77fe6e0d3a09a/H2B9T4dtpFxOgxTZ3RpkP.png" width="800">

✅ Responsible AI Considerations

Like other large language models, InfiFusion may exhibit behaviors that raise concerns around fairness, safety, and reliability. While our fusion framework enhances reasoning and relational inference, the foundation models it integrates—including the pivot model phi-4—carry inherited limitations. Users should be aware of the following considerations:

• Language Coverage & Bias: InfiFusion is primarily trained and evaluated on English datasets. Its performance on non-English inputs may be degraded. Moreover, any biases or stereotypes present in the underlying models or datasets may be preserved or amplified through the fusion process.

• Representation of Harms & Perpetuation of Stereotypes: The fused models may over- or under-represent certain groups or reinforce societal stereotypes. Although quality filters and alignment procedures are used, they cannot fully eliminate harmful representations due to real-world imbalances in the data sources.

• Content Safety: The model may generate inappropriate, offensive, or unsafe content, especially in unconstrained or adversarial inputs. It is not recommended for use in sensitive domains (e.g., mental health, legal advice) without additional safeguards.

• Fused Behavior Complexity: Due to the fusion of multiple distinct models, InfiFusion may exhibit complex or emergent behaviors not present in any single model. This makes interpretability and debugging more challenging, especially in high-risk applications.

• Factuality & Hallucination: Like other generative models, InfiFusion may produce convincing yet factually incorrect or outdated content. Developers should not rely on model output as a source of truth and are encouraged to integrate verification mechanisms such as Retrieval-Augmented Generation (RAG).

• Code Generation Limitations: While InfiFusion includes training on code-related datasets, it may still produce invalid, unsafe, or incomplete code snippets. Outputs involving unfamiliar libraries or languages should be carefully reviewed before use.

---

Developer Responsibilities

Developers using InfiFusion are encouraged to:

• Evaluate outputs contextually, particularly for fairness, accuracy, and safety.
• Follow all applicable laws and regulations, including those relating to privacy, trade compliance, and data use.
• Avoid deployment in high-stakes decision-making (e.g., employment, finance, law enforcement) without extensive validation and domain-specific safeguards.
• Clearly disclose to users that they are interacting with an AI system, following transparency and responsible AI best practices.

By using this model, you agree to evaluate and manage risks responsibly and ensure your applications align with ethical and regulatory expectations.

🗒️ BibTex Citation

If you find this work helpful, feel free to give us a cite.

@article{yan2025infifusion,
  title={InfiFusion: A Unified Framework for Enhanced Cross-Model Reasoning via LLM Fusion},
  author={Zhaoyi Yan, Yiming Zhang, Baoyi He, Yuhao Fu, Qi Zhou, Zhijie Sang, Chunlin Ji, Shengyu Zhang, Fei Wu, Hongxia Yang},
  journal={arXiv preprint arXiv:2501.02795},
  year={2025}
}