Today's Fine-tuning & Training: Fastest-Growing Projects — April 22, 2026
Today's the Fine-tuning & Training space, we're seeing a surge of interest in multimodal models and efficient training methods. Repositories leveraging PyTorch and Metal Performance Shaders are gaining traction, while others focus on injecting high-capacity conditional memory into large language models (LLMs). The trend suggests a growing need for more effective and efficient fine-tuning techniques.
mattmireles/gemma-tuner-multimodal takes the top spot with a growth score of 75.13 and 1,366 stars. This repository allows users to fine-tune Gemma 4 and 3n with audio, images, and text on Apple Silicon using PyTorch and Metal Performance Shaders. Its popularity can be attributed to its innovative approach to multimodal learning and the increasing demand for more effective training methods.
facebookresearch/tribev2 comes in second with a growth score of 56.24 and an impressive 1,933 stars. This repository contains code to train and evaluate TRIBE v2, a multimodal model for brain response prediction. Its growing popularity likely stems from its cutting-edge approach to multimodal learning and the potential applications in fields like neuroscience.
QingGo/engram-peft has seen significant growth with a score of 31.90 and 29 stars. This unofficial implementation of DeepSeek Engram injects high-capacity conditional memory into LLMs via sparse retrieval PEFT without increasing inference FLOPs. Its growing popularity can be attributed to its innovative approach to efficient training methods and the potential applications in natural language processing.
0xSero/turboquant boasts a growth score of 27.96 and 1,146 stars. TurboQuant offers near-optimal KV cache quantization for LLM inference with Triton kernels + vLLM integration. Its popularity likely stems from its focus on efficient training methods and the growing need for more effective LLM optimization techniques.
ZJU-OmniAI/GFT has seen notable growth with a score of 24.58 and 25 stars. This repository focuses on reward fine-tuning with unbiased group advantages and dynamic coefficient rectification. Its popularity can be attributed to its innovative approach to reinforcement learning and the potential applications in fields like robotics.
tonbistudio/turboquant-pytorch has a growth score of 24.48 and an impressive 951 stars. This from-scratch PyTorch implementation of Google's TurboQuant offers 5x compression at 3-bit with 99.5% attention fidelity. Its popularity likely stems from its focus on efficient training methods and the growing need for more effective LLM optimization techniques.
WillowHe/EvoOpt_oppangu_optimization_model has seen growth with a score of 11.20 and 494 stars. This repository provides solutions leveraging Openpangu - 7B as the base model for fine-tuning and application of large language models (LLMs) in operations research optimization tasks. Its popularity can be attributed to its focus on practical applications of LLMs.
SUM-INNOVATION/RUMUS boasts a growth score of 9.45 and 116 stars. This Rust-based framework allows users to train neural networks efficiently. Its growing popularity likely stems from the increasing demand for more efficient training methods and the potential applications in fields like computer vision.
semidark/kokoro-deutsch has seen notable growth with a score of 8.88 and 24 stars. This project provides a complete, documented training recipe for fine-tuning Kokoro-82M on German. Its popularity can be attributed to its focus on practical applications of LLMs and the growing need for more effective language models.
verl-project/bumblebee rounds out the list with a growth score of 6.71 and 64 stars. This lightweight distributed training library offers a runtime API for orchestration, composable primitives for implementation work, and model composition plus registration hooks. Its popularity likely stems from its focus on efficient training methods and the growing need for more effective LLM optimization techniques.
mattmireles/gemma-tuner-multimodal takes the top spot with a growth score of 75.13 and 1,366 stars. This repository allows users to fine-tune Gemma 4 and 3n with audio, images, and text on Apple Silicon using PyTorch and Metal Performance Shaders. Its popularity can be attributed to its innovative approach to multimodal learning and the increasing demand for more effective training methods.
facebookresearch/tribev2 comes in second with a growth score of 56.24 and an impressive 1,933 stars. This repository contains code to train and evaluate TRIBE v2, a multimodal model for brain response prediction. Its growing popularity likely stems from its cutting-edge approach to multimodal learning and the potential applications in fields like neuroscience.
QingGo/engram-peft has seen significant growth with a score of 31.90 and 29 stars. This unofficial implementation of DeepSeek Engram injects high-capacity conditional memory into LLMs via sparse retrieval PEFT without increasing inference FLOPs. Its growing popularity can be attributed to its innovative approach to efficient training methods and the potential applications in natural language processing.
0xSero/turboquant boasts a growth score of 27.96 and 1,146 stars. TurboQuant offers near-optimal KV cache quantization for LLM inference with Triton kernels + vLLM integration. Its popularity likely stems from its focus on efficient training methods and the growing need for more effective LLM optimization techniques.
ZJU-OmniAI/GFT has seen notable growth with a score of 24.58 and 25 stars. This repository focuses on reward fine-tuning with unbiased group advantages and dynamic coefficient rectification. Its popularity can be attributed to its innovative approach to reinforcement learning and the potential applications in fields like robotics.
tonbistudio/turboquant-pytorch has a growth score of 24.48 and an impressive 951 stars. This from-scratch PyTorch implementation of Google's TurboQuant offers 5x compression at 3-bit with 99.5% attention fidelity. Its popularity likely stems from its focus on efficient training methods and the growing need for more effective LLM optimization techniques.
WillowHe/EvoOpt_oppangu_optimization_model has seen growth with a score of 11.20 and 494 stars. This repository provides solutions leveraging Openpangu - 7B as the base model for fine-tuning and application of large language models (LLMs) in operations research optimization tasks. Its popularity can be attributed to its focus on practical applications of LLMs.
SUM-INNOVATION/RUMUS boasts a growth score of 9.45 and 116 stars. This Rust-based framework allows users to train neural networks efficiently. Its growing popularity likely stems from the increasing demand for more efficient training methods and the potential applications in fields like computer vision.
semidark/kokoro-deutsch has seen notable growth with a score of 8.88 and 24 stars. This project provides a complete, documented training recipe for fine-tuning Kokoro-82M on German. Its popularity can be attributed to its focus on practical applications of LLMs and the growing need for more effective language models.
verl-project/bumblebee rounds out the list with a growth score of 6.71 and 64 stars. This lightweight distributed training library offers a runtime API for orchestration, composable primitives for implementation work, and model composition plus registration hooks. Its popularity likely stems from its focus on efficient training methods and the growing need for more effective LLM optimization techniques.