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 techniques. Repositories that provide solutions for fine-tuning large language models (LLMs) with various input types, such as audio and images, are gaining significant traction. Meanwhile, tools that optimize LLM inference through compression and caching are also attracting attention.
mattmireles/gemma-tuner-multimodal takes the top spot this week with a growth score of 80.25 and 1,359 stars. This repository allows users to fine-tune Gemma 4 and 3n models on Apple Silicon using PyTorch and Metal Performance Shaders, making it an attractive solution for multimodal tasks involving audio, images, and text. Its high growth rate can be attributed to the increasing demand for efficient multimodal training methods.
facebookresearch/tribev2 has a respectable growth score of 57.70 and 1,923 stars. This repository contains the code for TRIBE v2, a multimodal model designed for brain response prediction tasks. Although it only saw five commits in the past month, its established reputation and high star count suggest that researchers are interested in exploring its capabilities.
QingGo/engram-peft boasts an impressive 80 commits in the last 30 days and has garnered 28 stars. This unofficial implementation of DeepSeek Engram enables users to inject high-capacity conditional memory into LLMs via sparse retrieval PEFT, making it a valuable resource for those looking to optimize their models without increasing inference FLOPs.
0xSero/turboquant has a growth score of 28.46 and 1,126 stars. This repository provides TurboQuant, a near-optimal KV cache quantization technique for LLM inference that offers significant compression with minimal quality loss. Its moderate growth rate indicates ongoing interest in efficient training methods.
tonbistudio/turboquant-pytorch has a growth score of 25.37 and 950 stars. This from-scratch PyTorch implementation of TurboQuant is gaining traction due to its ability to provide 5x compression at 3-bit with 99.5% attention fidelity, making it an attractive solution for those seeking efficient LLM inference.
WillowHe/EvoOpt_oppangu_optimization_model has a growth score of 11.27 and 475 stars. This repository offers solutions for fine-tuning Openpangu-7B on operations research optimization tasks, demonstrating the growing interest in applying LLMs to specific domains.
semidark/kokoro-deutsch boasts 29 commits in the last month and has garnered 22 stars. This project provides a complete training recipe for fine-tuning Kokoro-82M on German language data, making it a valuable resource for those interested in exploring multilingual models.
OnlyTerp/turboquant has a growth score of 6.19 and 55 stars. As another implementation of TurboQuant, this repository offers near-optimal KV cache compression for LLM inference with minimal quality loss, further underscoring the demand for efficient training techniques.
Other notable mentions include verl-project/bumblebee, which provides a lightweight distributed training library for large language models, and SUM-INNOVATION/RUMUS, a Rust-based framework for training neural networks.
mattmireles/gemma-tuner-multimodal takes the top spot this week with a growth score of 80.25 and 1,359 stars. This repository allows users to fine-tune Gemma 4 and 3n models on Apple Silicon using PyTorch and Metal Performance Shaders, making it an attractive solution for multimodal tasks involving audio, images, and text. Its high growth rate can be attributed to the increasing demand for efficient multimodal training methods.
facebookresearch/tribev2 has a respectable growth score of 57.70 and 1,923 stars. This repository contains the code for TRIBE v2, a multimodal model designed for brain response prediction tasks. Although it only saw five commits in the past month, its established reputation and high star count suggest that researchers are interested in exploring its capabilities.
QingGo/engram-peft boasts an impressive 80 commits in the last 30 days and has garnered 28 stars. This unofficial implementation of DeepSeek Engram enables users to inject high-capacity conditional memory into LLMs via sparse retrieval PEFT, making it a valuable resource for those looking to optimize their models without increasing inference FLOPs.
0xSero/turboquant has a growth score of 28.46 and 1,126 stars. This repository provides TurboQuant, a near-optimal KV cache quantization technique for LLM inference that offers significant compression with minimal quality loss. Its moderate growth rate indicates ongoing interest in efficient training methods.
tonbistudio/turboquant-pytorch has a growth score of 25.37 and 950 stars. This from-scratch PyTorch implementation of TurboQuant is gaining traction due to its ability to provide 5x compression at 3-bit with 99.5% attention fidelity, making it an attractive solution for those seeking efficient LLM inference.
WillowHe/EvoOpt_oppangu_optimization_model has a growth score of 11.27 and 475 stars. This repository offers solutions for fine-tuning Openpangu-7B on operations research optimization tasks, demonstrating the growing interest in applying LLMs to specific domains.
semidark/kokoro-deutsch boasts 29 commits in the last month and has garnered 22 stars. This project provides a complete training recipe for fine-tuning Kokoro-82M on German language data, making it a valuable resource for those interested in exploring multilingual models.
OnlyTerp/turboquant has a growth score of 6.19 and 55 stars. As another implementation of TurboQuant, this repository offers near-optimal KV cache compression for LLM inference with minimal quality loss, further underscoring the demand for efficient training techniques.
Other notable mentions include verl-project/bumblebee, which provides a lightweight distributed training library for large language models, and SUM-INNOVATION/RUMUS, a Rust-based framework for training neural networks.