Today's Fine-tuning & Training: Fastest-Growing Projects — April 22, 2026
Today's Fine-tuning & Training space is marked by a surge in interest around multimodal models and efficient training methods. The top repositories are focusing on fine-tuning large language models (LLMs) with various inputs, such as audio, images, and text, to improve their performance. Meanwhile, other projects are exploring ways to optimize training processes, including quantization techniques and distributed training libraries.
mattmireles/gemma-tuner-multimodal takes the top spot with a growth score of 75.50 and 1,374 stars. This repository provides a way to fine-tune Gemma 4 and 3n models using PyTorch and Metal Performance Shaders on Apple Silicon, allowing for efficient processing of multimodal inputs like audio, images, and text. Its rapid growth can be attributed to the increasing demand for models that can handle diverse input types.
facebookresearch/tribev2 has a growth score of 56.71 and 1,945 stars. This repository contains code for training and evaluating TRIBE v2, a multimodal model designed for brain response prediction. Its popularity stems from its innovative approach to modeling complex cognitive processes, making it an attractive tool for researchers in the field.
QingGo/engram-peft boasts a growth score of 31.95 and 30 stars. This unofficial implementation of DeepSeek Engram injects high-capacity conditional memory into LLMs via sparse retrieval PEFT without increasing inference FLOPs. Its growing interest is likely due to its potential to enhance the performance of existing language models without sacrificing efficiency.
0xSero/turboquant has a growth score of 28.18 and 1,152 stars. TurboQuant offers near-optimal KV cache quantization for LLM inference using Triton kernels and vLLM integration. Its popularity can be attributed to its ability to significantly reduce the computational requirements of large language models during inference.
ZJU-OmniAI/GFT has a growth score of 24.67 and 26 stars. This repository introduces GFT, a method for fine-tuning with unbiased group advantages and dynamic coefficient rectification. Its growing interest is likely due to its potential to improve the efficiency and effectiveness of reinforcement learning algorithms.
tonbistudio/turboquant-pytorch boasts a growth score of 24.57 and 953 stars. This PyTorch implementation of Google's TurboQuant provides 5x compression at 3-bit with 99.5% attention fidelity for LLM KV cache compression. Its popularity stems from its ability to efficiently compress large language models while maintaining their performance.
WillowHe/EvoOpt_oppangu_optimization_model has a growth score of 11.63 and 514 stars. This repository provides solutions for fine-tuning Openpangu - 7B on operations research optimization tasks. Its growing interest is likely due to its potential applications in solving complex problems in fields like logistics and finance.
SUM-INNOVATION/RUMUS has a growth score of 9.73 and 125 stars. This Rust-based framework allows users to train neural networks efficiently. Its popularity can be attributed to the increasing demand for fast and efficient training methods.
semidark/kokoro-deutsch boasts a growth score of 9.00 and 24 stars. This project provides a complete recipe for fine-tuning Kokoro-82M on German. Its growing interest is likely due to its potential applications in natural language processing tasks specific to the German language.
verl-project/bumblebee has a growth score of 6.71 and 64 stars. Bumblebee is a lightweight distributed training library for large language models, exposing a runtime API for orchestration and composable primitives for implementation work. Its popularity stems from its potential to simplify the process of training complex models in parallel environments.
mattmireles/gemma-tuner-multimodal takes the top spot with a growth score of 75.50 and 1,374 stars. This repository provides a way to fine-tune Gemma 4 and 3n models using PyTorch and Metal Performance Shaders on Apple Silicon, allowing for efficient processing of multimodal inputs like audio, images, and text. Its rapid growth can be attributed to the increasing demand for models that can handle diverse input types.
facebookresearch/tribev2 has a growth score of 56.71 and 1,945 stars. This repository contains code for training and evaluating TRIBE v2, a multimodal model designed for brain response prediction. Its popularity stems from its innovative approach to modeling complex cognitive processes, making it an attractive tool for researchers in the field.
QingGo/engram-peft boasts a growth score of 31.95 and 30 stars. This unofficial implementation of DeepSeek Engram injects high-capacity conditional memory into LLMs via sparse retrieval PEFT without increasing inference FLOPs. Its growing interest is likely due to its potential to enhance the performance of existing language models without sacrificing efficiency.
0xSero/turboquant has a growth score of 28.18 and 1,152 stars. TurboQuant offers near-optimal KV cache quantization for LLM inference using Triton kernels and vLLM integration. Its popularity can be attributed to its ability to significantly reduce the computational requirements of large language models during inference.
ZJU-OmniAI/GFT has a growth score of 24.67 and 26 stars. This repository introduces GFT, a method for fine-tuning with unbiased group advantages and dynamic coefficient rectification. Its growing interest is likely due to its potential to improve the efficiency and effectiveness of reinforcement learning algorithms.
tonbistudio/turboquant-pytorch boasts a growth score of 24.57 and 953 stars. This PyTorch implementation of Google's TurboQuant provides 5x compression at 3-bit with 99.5% attention fidelity for LLM KV cache compression. Its popularity stems from its ability to efficiently compress large language models while maintaining their performance.
WillowHe/EvoOpt_oppangu_optimization_model has a growth score of 11.63 and 514 stars. This repository provides solutions for fine-tuning Openpangu - 7B on operations research optimization tasks. Its growing interest is likely due to its potential applications in solving complex problems in fields like logistics and finance.
SUM-INNOVATION/RUMUS has a growth score of 9.73 and 125 stars. This Rust-based framework allows users to train neural networks efficiently. Its popularity can be attributed to the increasing demand for fast and efficient training methods.
semidark/kokoro-deutsch boasts a growth score of 9.00 and 24 stars. This project provides a complete recipe for fine-tuning Kokoro-82M on German. Its growing interest is likely due to its potential applications in natural language processing tasks specific to the German language.
verl-project/bumblebee has a growth score of 6.71 and 64 stars. Bumblebee is a lightweight distributed training library for large language models, exposing a runtime API for orchestration and composable primitives for implementation work. Its popularity stems from its potential to simplify the process of training complex models in parallel environments.