Today's Fine-tuning & Training: Fastest-Growing Projects — April 20, 2026
Today's the Fine-tuning & Training space, we're seeing a surge in innovations around multimodal models and large language model (LLM) optimization. Researchers are pushing the boundaries of what's possible with AI, exploring new architectures and techniques to improve performance and efficiency. As a result, we're witnessing rapid growth in repositories that offer cutting-edge solutions for fine-tuning and training.
mattmireles/gemma-tuner-multimodal is leading the pack with an impressive Growth Score of 86.08 and 1,353 stars. This repository provides a framework for fine-tuning Gemma 4 and 3n models using PyTorch and Metal Performance Shaders, allowing researchers to leverage audio, images, and text on Apple Silicon. Its rapid growth can be attributed to its innovative approach to multimodal learning, making it an attractive solution for those looking to explore new frontiers in AI research.
facebookresearch/tribev2 boasts 1,917 stars and a Growth Score of 59.67, despite relatively low commit activity over the past month. This repository contains code for training and evaluating TRIBE v2, a multimodal model designed for brain response prediction. Its growth can be attributed to its significance in the field of neuroscience and AI research, as well as the reputation of Facebook Research.
QingGo/engram-peft has gained significant traction with 28 stars and a Growth Score of 32.69, driven by its innovative approach to injecting high-capacity conditional memory into LLMs via sparse retrieval PEFT. This unofficial implementation of DeepSeek Engram offers a promising solution for improving LLM performance without increasing inference FLOPs.
0xSero/turboquant has attracted 1,120 stars and boasts a Growth Score of 29.33, despite limited commit activity over the past month. This repository provides near-optimal KV cache quantization for LLM inference using Triton kernels and vLLM integration. Its growth can be attributed to its potential to significantly improve LLM performance while reducing computational costs.
tonbistudio/turboquant-pytorch offers a PyTorch implementation of Google's TurboQuant, achieving 5x compression at 3-bit with 99.5% attention fidelity. With 948 stars and a Growth Score of 26.31, this repository is gaining popularity among researchers looking for efficient LLM solutions.
WillowHe/EvoOpt_oppangu_optimization_model provides a set of solutions leveraging Openpangu-7B as the base model for fine-tuning and application of LLMs in operations research optimization tasks. With 442 stars and a Growth Score of 11.02, this repository is attracting attention from researchers interested in exploring new applications for LLMs.
SUM-INNOVATION/RUMUS offers a Rust-based framework for training neural networks, with 98 stars and a Growth Score of 9.65. Despite limited growth, this repository provides an interesting alternative to traditional frameworks, making it worth keeping an eye on.
verl-project/bumblebee is a lightweight distributed training library for large language models, boasting 61 stars and a Growth Score of 8.20. Its composable primitives and model composition capabilities make it an attractive solution for researchers looking to scale their LLM training efforts.
OnlyTerp/turboquant offers the first open-source implementation of Google TurboQuant, achieving near-optimal KV cache compression for LLM inference with 55 stars and a Growth Score of 6.42. Its growth can be attributed to its significance in the field of LLM research and optimization.
Dynamis-Labs/spectralquant proposes an innovative approach to breaking TurboQuant's compression limit via spectral structure, attracting 121 stars and a Growth Score of 6.03. This repository offers a promising solution for further improving LLM performance and efficiency.
mattmireles/gemma-tuner-multimodal is leading the pack with an impressive Growth Score of 86.08 and 1,353 stars. This repository provides a framework for fine-tuning Gemma 4 and 3n models using PyTorch and Metal Performance Shaders, allowing researchers to leverage audio, images, and text on Apple Silicon. Its rapid growth can be attributed to its innovative approach to multimodal learning, making it an attractive solution for those looking to explore new frontiers in AI research.
facebookresearch/tribev2 boasts 1,917 stars and a Growth Score of 59.67, despite relatively low commit activity over the past month. This repository contains code for training and evaluating TRIBE v2, a multimodal model designed for brain response prediction. Its growth can be attributed to its significance in the field of neuroscience and AI research, as well as the reputation of Facebook Research.
QingGo/engram-peft has gained significant traction with 28 stars and a Growth Score of 32.69, driven by its innovative approach to injecting high-capacity conditional memory into LLMs via sparse retrieval PEFT. This unofficial implementation of DeepSeek Engram offers a promising solution for improving LLM performance without increasing inference FLOPs.
0xSero/turboquant has attracted 1,120 stars and boasts a Growth Score of 29.33, despite limited commit activity over the past month. This repository provides near-optimal KV cache quantization for LLM inference using Triton kernels and vLLM integration. Its growth can be attributed to its potential to significantly improve LLM performance while reducing computational costs.
tonbistudio/turboquant-pytorch offers a PyTorch implementation of Google's TurboQuant, achieving 5x compression at 3-bit with 99.5% attention fidelity. With 948 stars and a Growth Score of 26.31, this repository is gaining popularity among researchers looking for efficient LLM solutions.
WillowHe/EvoOpt_oppangu_optimization_model provides a set of solutions leveraging Openpangu-7B as the base model for fine-tuning and application of LLMs in operations research optimization tasks. With 442 stars and a Growth Score of 11.02, this repository is attracting attention from researchers interested in exploring new applications for LLMs.
SUM-INNOVATION/RUMUS offers a Rust-based framework for training neural networks, with 98 stars and a Growth Score of 9.65. Despite limited growth, this repository provides an interesting alternative to traditional frameworks, making it worth keeping an eye on.
verl-project/bumblebee is a lightweight distributed training library for large language models, boasting 61 stars and a Growth Score of 8.20. Its composable primitives and model composition capabilities make it an attractive solution for researchers looking to scale their LLM training efforts.
OnlyTerp/turboquant offers the first open-source implementation of Google TurboQuant, achieving near-optimal KV cache compression for LLM inference with 55 stars and a Growth Score of 6.42. Its growth can be attributed to its significance in the field of LLM research and optimization.
Dynamis-Labs/spectralquant proposes an innovative approach to breaking TurboQuant's compression limit via spectral structure, attracting 121 stars and a Growth Score of 6.03. This repository offers a promising solution for further improving LLM performance and efficiency.