Hugging Face's logo

Spaces:
InfiX-ai
/
README
Running

App Files Community
README / README.md
congkai's picture
congkai
Update README.md
fb40b16 verified
preview code
|
raw
history blame contribute delete
5.51 kB
metadata
title: README
emoji: 🏆
colorFrom: red
colorTo: yellow
sdk: static
pinned: false

InfiX.ai

Welcome to InfiX.ai! We believe our research will eventually lead to decentralized Generative AI—a future where everyone can access, contribute to, and benefit from AI equally.
Our Mission: Generative AI for all, intelligence in every task.

🤖 Our Model Series

🔗 Model Fusion & Model Merging

Model fusion refers to the process of combining multiple trained models—often from different domains, architectures, or training datasets—into a single, more powerful model. The goal is to integrate their strengths and knowledge, improving performance, generalization, or efficiency.
Model merging is a specific type of model fusion that involves combining the internal parameters (typically weights) of two or more pretrained models to produce a single model that inherits knowledge from all sources. Unlike ensemble methods, model merging produces a single merged model rather than relying on multiple models at inference time.

  • InfiFusion: InfiFusion is a logit-level fusion pipeline based on Universal Logit Distillation, enhanced with Top-K filtering and logits standardization. It supports both pairwise and unified fusion strategies to balance performance and efficiency.
  • InfiGFusion: InfiGFusion is a structure-aware extension that builds co-activation graphs from logits and aligns them via an efficient Gromov-Wasserstein loss approximation, capturing cross-dimension semantic dependencies for stronger reasoning.
  • InfiFPO: InfiFPO is a lightweight fusion method during the preference alignment phase that injects fused model behavior into preference learning, enabling richer signal during DPO-style fine-tuning.

🧠 Reasoning-Enhanced Low-Resource Training Pipeline

  • InfiR: InfiR aims to advance AI systems by improving reasoning, reducing adoption barriers, and addressing privacy concerns through smaller model sizes.
  • InfiR-FP8: InfiR-FP8 is a smaller reasoning-enhanced model trained from scratch using FP8 precision, achieving successful convergence while reducing memory usage by 10% and improving training speed by 20% during the training process. The model will be released in mid-September.
  • InfiAlign: InfiAlign is a scalable and data-efficient post-training framework that combines supervised fine-tuning (SFT) and reinforcement learning (RL) with a high-quality data selection pipeline to enhance reasoning in large language models.
  • InfiMMR: InfiMMR is a novel three-phase curriculum framework that systematically enhances multimodal reasoning capabilities in small language models through foundational reasoning activation, cross-modal adaptation, and multimodal reasoning enhancement.

🖥️ Advanced Vision-Native Agent for GUI Interaction

  • InfiGUIAgent: InfiGUIAgent is a GUI agent that embeds native hierarchical and expectation-reflection reasoning through a unique two-stage supervised pipeline, enabling robust, multi-step GUI task automation.
  • InfiGUI-R1: InfiGUI-R1 is a GUI agent developed via the Actor2Reasoner framework, which evolves a reactive model into a deliberative reasoner capable of sophisticated planning and error recovery through spatial reasoning distillation and reinforcement learning.
  • InfiGUI-G1: InfiGUI-G1 is a multimodal GUI agent that employs Adaptive Exploration Policy Optimization (AEPO) to improve semantic alignment in GUI grounding. The novel training framework achieves up to 8.3% relative improvement over baseline methods.

📰 News