Today's Fine-tuning & Training: Fastest-Growing Projects — April 19, 2026
Today's Fine-tuning & Training, we're seeing a surge of interest in multimodal models and innovative techniques for compressing large language model (LLM) caches. The top-growing repositories are showcasing cutting-edge approaches to fine-tuning and training AI models, with a focus on improving efficiency and performance.
The fastest-growing repository this week is mattmireles/gemma-tuner-multimodal, which boasts an impressive growth score of 92.50 and 1,338 stars. This project enables the fine-tuning of Gemma 4 and 3n with audio, images, and text on Apple Silicon using PyTorch and Metal Performance Shaders. Its rapid growth is likely due to its innovative approach to multimodal learning and its potential applications in areas like computer vision and natural language processing.
Another notable repository is facebookresearch/tribev2, which has a growth score of 61.44 and 1,902 stars. This project provides the code for training and evaluating TRIBE v2, a multimodal model for brain response prediction. Its growth can be attributed to its cutting-edge approach to multimodal learning and its potential applications in areas like neuroscience and cognitive psychology.
QingGo/engram-peft is also gaining traction, with a growth score of 37.14 and 25 stars. This project implements DeepSeek Engram, which injects high-capacity conditional memory into LLMs via sparse retrieval PEFT without increasing inference FLOPs. Its growth is likely due to its innovative approach to improving the efficiency of LLMs.
TurboQuant, a near-optimal KV cache quantization technique for LLM inference, is another popular topic this week. 0xSero/turboquant has a growth score of 29.72 and 1,093 stars, while tonbistudio/turboquant-pytorch has a growth score of 27.20 and 943 stars. Both projects provide implementations of TurboQuant, with the latter being a from-scratch PyTorch implementation that achieves 5x compression at 3-bit with 99.5% attention fidelity.
Other notable repositories include WillowHe/EvoOpt_oppangu_optimization_model, which provides solutions for fine-tuning and applying LLMs to operations research optimization tasks; SUM-INNOVATION/RUMUS, a Rust-based framework for training neural networks; and Dynamis-Labs/spectralquant, which breaks TurboQuant's compression limit via spectral structure.
The fastest-growing repository this week is mattmireles/gemma-tuner-multimodal, which boasts an impressive growth score of 92.50 and 1,338 stars. This project enables the fine-tuning of Gemma 4 and 3n with audio, images, and text on Apple Silicon using PyTorch and Metal Performance Shaders. Its rapid growth is likely due to its innovative approach to multimodal learning and its potential applications in areas like computer vision and natural language processing.
Another notable repository is facebookresearch/tribev2, which has a growth score of 61.44 and 1,902 stars. This project provides the code for training and evaluating TRIBE v2, a multimodal model for brain response prediction. Its growth can be attributed to its cutting-edge approach to multimodal learning and its potential applications in areas like neuroscience and cognitive psychology.
QingGo/engram-peft is also gaining traction, with a growth score of 37.14 and 25 stars. This project implements DeepSeek Engram, which injects high-capacity conditional memory into LLMs via sparse retrieval PEFT without increasing inference FLOPs. Its growth is likely due to its innovative approach to improving the efficiency of LLMs.
TurboQuant, a near-optimal KV cache quantization technique for LLM inference, is another popular topic this week. 0xSero/turboquant has a growth score of 29.72 and 1,093 stars, while tonbistudio/turboquant-pytorch has a growth score of 27.20 and 943 stars. Both projects provide implementations of TurboQuant, with the latter being a from-scratch PyTorch implementation that achieves 5x compression at 3-bit with 99.5% attention fidelity.
Other notable repositories include WillowHe/EvoOpt_oppangu_optimization_model, which provides solutions for fine-tuning and applying LLMs to operations research optimization tasks; SUM-INNOVATION/RUMUS, a Rust-based framework for training neural networks; and Dynamis-Labs/spectralquant, which breaks TurboQuant's compression limit via spectral structure.