About InfiX.ai
We are a research team dedicated to advancing AI through model fusion, GUI agents, and multimodal intelligence.
Our Mission
At InfiX.ai, we believe in making AI more accessible, efficient, and capable. Our research focuses on developing innovative techniques that push the boundaries of what AI systems can achieve while maintaining practical applicability.
We are committed to open science and share our models, datasets, and research findings with the community to accelerate progress in AI.
Research Focus
Our research focuses on two primary areas that are reshaping how AI models are developed and deployed:
Model Fusion & Model Merging
We pioneer advanced techniques for combining AI models to create more powerful and efficient systems:
- Model Merging: Combining homogeneous models with the same architecture directly in parameter space to produce a single checkpoint with baseline-like inference cost.
- Model Fusion: Combining heterogeneous or homogeneous models in prediction/knowledge space through ensembles, logit averaging, and distillation.
Reasoning-Enhanced Low-Resource Training
We develop methods to create highly capable AI systems that require minimal computational resources:
- Edge AI Deployment: Running sophisticated models on resource-constrained devices
- Privacy-Preserving AI: Local processing without cloud dependencies
- Cost-Effective Solutions: Reducing computational and infrastructure costs
Our Research Innovations
InfiFusion Series
- InfiFusion: Logit-level fusion pipeline based on Universal Logit Distillation with Top-K filtering and logits standardization.
- InfiGFusion: Structure-aware extension using co-activation graphs and Gromov-Wasserstein loss for stronger reasoning.
- InfiPPO: Lightweight fusion during preference alignment phase for richer signal in DPO-style fine-tuning.
InfiR: Reasoning-Enhanced Training
InfiR advances AI systems by improving reasoning capabilities and reducing adoption barriers through smaller model sizes with FP8 precision training for enhanced efficiency.
- • Democratized access for smaller organizations
- • Ultra-efficient FP8 precision training
- • Graph-based reasoning methods
GUI Agents
Creating intelligent agents capable of understanding and interacting with graphical user interfaces across platforms, enabling automated workflows and accessibility solutions.
Multimodal AI
Building systems that can process and reason across multiple modalities including text, images, and user interfaces for comprehensive understanding.
Connect With Us
We welcome collaboration and engagement from the research community. Find our work on GitHub and HuggingFace.