Today's Fine-tuning & Training: Fastest-Growing Projects — April 19, 2026
This week, the Fine-tuning & Training space on GitHub saw significant activity around multimodal models and efficient training techniques. Several repositories focused on fine-tuning large language models (LLMs) for specific tasks, such as brain response prediction and operations research optimization. Meanwhile, others explored novel approaches to compression and quantization for LLMs.
mattmireles/gemma-tuner-multimodal leads the pack with a growth score of 92.67 and 1,342 stars. This repository provides a fine-tuning framework for Gemma 4 and 3n models using PyTorch and Metal Performance Shaders on Apple Silicon, allowing for multimodal training with audio, images, and text inputs. Its high growth rate can be attributed to the increasing interest in multimodal learning and the need for efficient training techniques.
facebookresearch/tribev2 boasts an impressive 1,907 stars and a growth score of 61.60. This repository contains the code for TRIBE v2, a multimodal model for brain response prediction, which is likely attracting attention from researchers working on neuroscience-related projects. Although its commit activity has been relatively low over the past month, its high star count indicates sustained interest in the project.
QingGo/engram-peft has gained 25 stars and achieved a growth score of 37.14, despite being an unofficial implementation of DeepSeek Engram. This repository provides a method for injecting high-capacity conditional memory into LLMs via sparse retrieval PEFT, which is likely appealing to developers looking for efficient ways to enhance their models. Its high commit activity over the past month (80 commits) suggests active development and community engagement.
0xSero/turboquant has garnered 1,099 stars and a growth score of 29.96. This repository focuses on near-optimal KV cache quantization for LLM inference using Triton kernels and vLLM integration. Its relatively low commit activity (2 commits over the past month) might indicate that the project is nearing completion or transitioning to maintenance mode.
tonbistudio/turboquant-pytorch has attracted 945 stars and achieved a growth score of 27.24. This repository provides a PyTorch implementation of Google's TurboQuant, offering 5x compression at 3-bit with high attention fidelity. Its moderate commit activity (6 commits over the past month) suggests ongoing development and refinement.
WillowHe/EvoOpt_oppangu_optimization_model has gained 433 stars and a growth score of 11.35. This repository offers solutions for fine-tuning Openpangu - 7B on operations research optimization tasks, which is likely appealing to researchers working in this domain. Although its commit activity has been relatively low (3 commits over the past month), its moderate star count indicates interest in the project.
SUM-INNOVATION/RUMUS boasts a growth score of 9.05 and 83 stars. This Rust-based framework for training neural networks is likely attracting attention from developers interested in alternative programming languages for machine learning tasks. Its high commit activity over the past month (35 commits) suggests active development and community engagement.
OnlyTerp/turboquant has gained 55 stars and achieved a growth score of 6.68. This repository provides an open-source implementation of Google TurboQuant, offering near-optimal KV cache compression for LLM inference. Its high commit activity over the past month (43 commits) suggests active development and refinement.
Mintzs/oogaboogalm has attracted 40 stars and a growth score of 6.39. This repository explores fine-tuning AI models to reduce token use by incorporating caveman system prompts and skills directly into the model itself. Its moderate commit activity over the past month (12 commits) suggests ongoing experimentation and development.
Dynamis-Labs/spectralquant boasts a growth score of 6.18 and 113 stars. This repository presents an alternative approach to compression, breaking TurboQuant's limit via spectral structure. Although its commit activity has been relatively low (3 commits over the past month), its moderate star count indicates interest in this novel approach.
mattmireles/gemma-tuner-multimodal leads the pack with a growth score of 92.67 and 1,342 stars. This repository provides a fine-tuning framework for Gemma 4 and 3n models using PyTorch and Metal Performance Shaders on Apple Silicon, allowing for multimodal training with audio, images, and text inputs. Its high growth rate can be attributed to the increasing interest in multimodal learning and the need for efficient training techniques.
facebookresearch/tribev2 boasts an impressive 1,907 stars and a growth score of 61.60. This repository contains the code for TRIBE v2, a multimodal model for brain response prediction, which is likely attracting attention from researchers working on neuroscience-related projects. Although its commit activity has been relatively low over the past month, its high star count indicates sustained interest in the project.
QingGo/engram-peft has gained 25 stars and achieved a growth score of 37.14, despite being an unofficial implementation of DeepSeek Engram. This repository provides a method for injecting high-capacity conditional memory into LLMs via sparse retrieval PEFT, which is likely appealing to developers looking for efficient ways to enhance their models. Its high commit activity over the past month (80 commits) suggests active development and community engagement.
0xSero/turboquant has garnered 1,099 stars and a growth score of 29.96. This repository focuses on near-optimal KV cache quantization for LLM inference using Triton kernels and vLLM integration. Its relatively low commit activity (2 commits over the past month) might indicate that the project is nearing completion or transitioning to maintenance mode.
tonbistudio/turboquant-pytorch has attracted 945 stars and achieved a growth score of 27.24. This repository provides a PyTorch implementation of Google's TurboQuant, offering 5x compression at 3-bit with high attention fidelity. Its moderate commit activity (6 commits over the past month) suggests ongoing development and refinement.
WillowHe/EvoOpt_oppangu_optimization_model has gained 433 stars and a growth score of 11.35. This repository offers solutions for fine-tuning Openpangu - 7B on operations research optimization tasks, which is likely appealing to researchers working in this domain. Although its commit activity has been relatively low (3 commits over the past month), its moderate star count indicates interest in the project.
SUM-INNOVATION/RUMUS boasts a growth score of 9.05 and 83 stars. This Rust-based framework for training neural networks is likely attracting attention from developers interested in alternative programming languages for machine learning tasks. Its high commit activity over the past month (35 commits) suggests active development and community engagement.
OnlyTerp/turboquant has gained 55 stars and achieved a growth score of 6.68. This repository provides an open-source implementation of Google TurboQuant, offering near-optimal KV cache compression for LLM inference. Its high commit activity over the past month (43 commits) suggests active development and refinement.
Mintzs/oogaboogalm has attracted 40 stars and a growth score of 6.39. This repository explores fine-tuning AI models to reduce token use by incorporating caveman system prompts and skills directly into the model itself. Its moderate commit activity over the past month (12 commits) suggests ongoing experimentation and development.
Dynamis-Labs/spectralquant boasts a growth score of 6.18 and 113 stars. This repository presents an alternative approach to compression, breaking TurboQuant's limit via spectral structure. Although its commit activity has been relatively low (3 commits over the past month), its moderate star count indicates interest in this novel approach.