Today's Fine-tuning & Training: Fastest-Growing Projects — April 23, 2026
Today's the Fine-tuning & Training space, we're seeing a surge in interest around tools that optimize and compress large language models (LLMs) for more efficient inference. Repositories focused on fine-tuning and training LLMs with multimodal inputs, sparse retrieval, and novel compression techniques are gaining significant traction.
mattmireles/gemma-tuner-multimodal is leading the pack with a growth score of 71.00 and 1,378 stars. This repository provides a PyTorch-based implementation for fine-tuning Gemma 4 and 3n models on Apple Silicon using Metal Performance Shaders, allowing for efficient training with audio, images, and text inputs. Its rapid growth can be attributed to the increasing demand for multimodal LLMs that can handle diverse input types.
QingGo/engram-peft is another notable repository, boasting a growth score of 29.09 and 31 stars. This unofficial implementation of DeepSeek Engram injects high-capacity conditional memory into LLMs via sparse retrieval PEFT, allowing for efficient inference without increasing FLOPs. Its growing popularity stems from the need for more efficient and scalable LLM architectures.
0xSero/turboquant has garnered significant attention with 1,168 stars and a growth score of 27.59. This repository provides a near-optimal KV cache quantization technique for LLM inference using Triton kernels and vLLM integration, achieving 3-bit keys and 2-bit values. Its growth can be attributed to the increasing demand for optimized LLM inference techniques.
tonbistudio/turboquant-pytorch is another implementation of Google's TurboQuant, with a growth score of 23.83 and 956 stars. This PyTorch-based repository achieves 5x compression at 3-bit with 99.5% attention fidelity, making it an attractive solution for LLM developers seeking efficient inference techniques.
ZJU-OmniAI/GFT is gaining traction with a growth score of 21.43 and 27 stars. This repository introduces a novel fine-tuning approach using unbiased group advantages and dynamic coefficient rectification, providing a more robust training framework for LLMs. Its growing popularity stems from the need for more effective fine-tuning techniques.
WillowHe/EvoOpt_oppangu_optimization_model offers a set of solutions leveraging Openpangu-7B as a base model for fine-tuning and optimization tasks in operations research, with a growth score of 11.15 and 514 stars. This repository's growth can be attributed to the increasing demand for LLMs in specialized domains.
SUM-INNOVATION/RUMUS is a Rust-based framework for training neural networks, boasting a growth score of 9.70 and 134 stars. Its growing popularity stems from the need for more efficient and scalable deep learning frameworks.
semidark/kokoro-deutsch provides a complete training recipe for fine-tuning Kokoro-82M on German language data, with a growth score of 8.92 and 25 stars. This repository's growth can be attributed to the increasing demand for multilingual LLMs.
verl-project/bumblebee is a lightweight distributed training library for large language models, with a growth score of 5.94 and 65 stars. Its growing popularity stems from the need for more efficient and scalable training frameworks.
OnlyTerp/turboquant offers another implementation of Google's TurboQuant, with a growth score of 5.84 and 57 stars. This repository achieves near-optimal KV cache compression for LLM inference, making it an attractive solution for developers seeking optimized techniques.
mattmireles/gemma-tuner-multimodal is leading the pack with a growth score of 71.00 and 1,378 stars. This repository provides a PyTorch-based implementation for fine-tuning Gemma 4 and 3n models on Apple Silicon using Metal Performance Shaders, allowing for efficient training with audio, images, and text inputs. Its rapid growth can be attributed to the increasing demand for multimodal LLMs that can handle diverse input types.
QingGo/engram-peft is another notable repository, boasting a growth score of 29.09 and 31 stars. This unofficial implementation of DeepSeek Engram injects high-capacity conditional memory into LLMs via sparse retrieval PEFT, allowing for efficient inference without increasing FLOPs. Its growing popularity stems from the need for more efficient and scalable LLM architectures.
0xSero/turboquant has garnered significant attention with 1,168 stars and a growth score of 27.59. This repository provides a near-optimal KV cache quantization technique for LLM inference using Triton kernels and vLLM integration, achieving 3-bit keys and 2-bit values. Its growth can be attributed to the increasing demand for optimized LLM inference techniques.
tonbistudio/turboquant-pytorch is another implementation of Google's TurboQuant, with a growth score of 23.83 and 956 stars. This PyTorch-based repository achieves 5x compression at 3-bit with 99.5% attention fidelity, making it an attractive solution for LLM developers seeking efficient inference techniques.
ZJU-OmniAI/GFT is gaining traction with a growth score of 21.43 and 27 stars. This repository introduces a novel fine-tuning approach using unbiased group advantages and dynamic coefficient rectification, providing a more robust training framework for LLMs. Its growing popularity stems from the need for more effective fine-tuning techniques.
WillowHe/EvoOpt_oppangu_optimization_model offers a set of solutions leveraging Openpangu-7B as a base model for fine-tuning and optimization tasks in operations research, with a growth score of 11.15 and 514 stars. This repository's growth can be attributed to the increasing demand for LLMs in specialized domains.
SUM-INNOVATION/RUMUS is a Rust-based framework for training neural networks, boasting a growth score of 9.70 and 134 stars. Its growing popularity stems from the need for more efficient and scalable deep learning frameworks.
semidark/kokoro-deutsch provides a complete training recipe for fine-tuning Kokoro-82M on German language data, with a growth score of 8.92 and 25 stars. This repository's growth can be attributed to the increasing demand for multilingual LLMs.
verl-project/bumblebee is a lightweight distributed training library for large language models, with a growth score of 5.94 and 65 stars. Its growing popularity stems from the need for more efficient and scalable training frameworks.
OnlyTerp/turboquant offers another implementation of Google's TurboQuant, with a growth score of 5.84 and 57 stars. This repository achieves near-optimal KV cache compression for LLM inference, making it an attractive solution for developers seeking optimized techniques.