Today's Fine-tuning & Training: Fastest-Growing Projects — April 21, 2026
Today's the Fine-tuning & Training space, we're seeing a surge in interest around multimodal models and efficient training methods. Repositories focusing on fine-tuning large language models (LLMs) with various data types, such as audio, images, and text, are gaining significant traction. Meanwhile, projects optimizing training processes for LLMs using techniques like quantization and sparse retrieval are also attracting attention.
mattmireles/gemma-tuner-multimodal is a standout repository this week, boasting an impressive growth score of 80.36 and 1,362 stars. This project allows users to fine-tune Gemma 4 and 3n models with audio, images, and text on Apple Silicon using PyTorch and Metal Performance Shaders, making it an attractive solution for those looking to experiment with multimodal inputs.
facebookresearch/tribev2 has a growth score of 58.04 and 1,927 stars, indicating its popularity among researchers. This repository contains the code to train and evaluate TRIBE v2, a multimodal model for brain response prediction, showcasing the growing interest in applying AI to complex tasks like neuroscience.
ZJU-OmniAI/GFT has a growth score of 29.30 and 23 stars, with 45 commits in the past 30 days. This project focuses on reward fine-tuning using unbiased group advantages and dynamic coefficient rectification, offering an innovative approach to training AI models. Its growth suggests that researchers are eager to explore new methods for improving model performance.
QingGo/engram-peft boasts a growth score of 29.11 and 29 stars, with 80 commits in the past month. This unofficial implementation of DeepSeek Engram enables users to inject high-capacity conditional memory into LLMs via sparse retrieval PEFT without increasing inference FLOPs, making it an attractive solution for those looking to improve model efficiency.
0xSero/turboquant has a growth score of 28.85 and 1,141 stars, with only 2 commits in the past 30 days. This project focuses on near-optimal KV cache quantization for LLM inference using Triton kernels and vLLM integration, demonstrating the growing interest in optimizing training processes for large models.
tonbistudio/turboquant-pytorch has a growth score of 25.39 and 951 stars, with 6 commits in the past month. This from-scratch PyTorch implementation of Google's TurboQuant offers 5x compression at 3-bit with 99.5% attention fidelity, making it an attractive solution for those looking to reduce model size without sacrificing performance.
The remaining repositories on this list also demonstrate growth and interest in fine-tuning and training AI models, including WillowHe/EvoOpt_oppangu_optimization_model, SUM-INNOVATION/RUMUS, semidark/kokoro-deutsch, and verl-project/bumblebee.
mattmireles/gemma-tuner-multimodal is a standout repository this week, boasting an impressive growth score of 80.36 and 1,362 stars. This project allows users to fine-tune Gemma 4 and 3n models with audio, images, and text on Apple Silicon using PyTorch and Metal Performance Shaders, making it an attractive solution for those looking to experiment with multimodal inputs.
facebookresearch/tribev2 has a growth score of 58.04 and 1,927 stars, indicating its popularity among researchers. This repository contains the code to train and evaluate TRIBE v2, a multimodal model for brain response prediction, showcasing the growing interest in applying AI to complex tasks like neuroscience.
ZJU-OmniAI/GFT has a growth score of 29.30 and 23 stars, with 45 commits in the past 30 days. This project focuses on reward fine-tuning using unbiased group advantages and dynamic coefficient rectification, offering an innovative approach to training AI models. Its growth suggests that researchers are eager to explore new methods for improving model performance.
QingGo/engram-peft boasts a growth score of 29.11 and 29 stars, with 80 commits in the past month. This unofficial implementation of DeepSeek Engram enables users to inject high-capacity conditional memory into LLMs via sparse retrieval PEFT without increasing inference FLOPs, making it an attractive solution for those looking to improve model efficiency.
0xSero/turboquant has a growth score of 28.85 and 1,141 stars, with only 2 commits in the past 30 days. This project focuses on near-optimal KV cache quantization for LLM inference using Triton kernels and vLLM integration, demonstrating the growing interest in optimizing training processes for large models.
tonbistudio/turboquant-pytorch has a growth score of 25.39 and 951 stars, with 6 commits in the past month. This from-scratch PyTorch implementation of Google's TurboQuant offers 5x compression at 3-bit with 99.5% attention fidelity, making it an attractive solution for those looking to reduce model size without sacrificing performance.
The remaining repositories on this list also demonstrate growth and interest in fine-tuning and training AI models, including WillowHe/EvoOpt_oppangu_optimization_model, SUM-INNOVATION/RUMUS, semidark/kokoro-deutsch, and verl-project/bumblebee.