Models

A quantization-aware distillation method that shrinks an 8B vision-language model into a 2B student and deploys it as genuine packed INT4 weights, retaining most of the teacher's accuracy.

• Models

A systems framework that coordinates GPU and NPU clusters for agentic reinforcement-learning post-training, routing requests across heterogeneous vLLM workers and adjusting resource allocation as workload changes.

• Models
• Inference

An autonomous research loop that has an AI agent iterate on a fixed vision-language model's training code, evaluate each change against the official DocVQA benchmark, and keep only changes that produce a real measured improvement.

• Models
• AI evals

A collection of end-to-end SageMaker notebooks and config-driven recipes covering generative-AI model training, fine-tuning, inference optimization, MLOps automation, distributed training, RAG, and agent-based applications.

• Models
• Tutorials

Research code for two visual tokenization methods, BSQ-ViT and its successor L24SQ-ViT, that use spherical vector quantization instead of a learned codebook to tokenize images and video.

• Models

A Graph Neural Network library built on Apple's MLX framework, designed to train and run GNNs on Apple Silicon's unified CPU and GPU memory, with reported speedups over other frameworks on large graphs.

• Models
• Libraries

A pure MLX reimplementation of the YOLO26 object detection model for Apple Silicon, with no PyTorch dependency at runtime, covering detection, tracking, segmentation, and pose.

• Models

A local FastAPI and React studio for editing a language model's identity and behavior by manipulating token directions revealed by a Jacobian lens, then exporting the edits into a standalone checkpoint that runs anywhere transformers models do.

• Models

An adaptive mixed-precision quantization format and runtime for Apple Silicon MLX, positioned as a GGUF equivalent that keeps large models quantized in GPU memory while running at full Metal speed.

• Models
• Inference

A lightweight, open-source vision-language model, about 1.2 billion parameters, for unified document parsing across both digital documents and camera-captured photos of documents.

• RAG
• Models

A memory-optimized ROCm build of Qwen 3.8 27B for AMD Strix Halo APUs, combining 4-bit block quantization, speculative decoding, and a KV-cache scheme to raise inference throughput on unified-memory hardware.

• Models
• Inference

A toolkit for training and evaluating German-language word embeddings, covering obtaining and preprocessing German text corpora, training models with gensim's word2vec, and evaluating them against generated test sets.

• Models

A repository combining a custom experimental machine learning algorithm for tabular data with a hand-curated, machine-readable knowledge base of roughly 900 university courses, research papers and explainers, each normalized into Markdown with full provenance.

• Models
• Curated lists

A CUDA and C++ implementation of quantized large language model training, aimed at single-node, multi-GPU training setups, inspired by llm.c.

• Models

A reinforcement learning framework built on Apple's MLX array framework, with single-file implementations of algorithms such as DQN, PPO and SAC that run end-to-end on device, including the environments, for over a million steps per second on Apple silicon.

• Models

A native macOS application for fine-tuning large language models on Apple Silicon, letting a user pick a model and a training algorithm and watch the training loss fall entirely on-device, with no cloud step required.

• Models

A fine-tuning skill for Claude Code, Gemini CLI and other agent hosts that runs the whole LoRA fine-tuning lifecycle from a plain-language request: environment setup, data formatting, training, evaluation and export.

• Agent skills
• Models

A small proof-of-concept language model, explicitly not built as an LLM, that predicts in a latent space, keeps internal recurrent state, and generates output byte by byte instead of token by token.

• Models

A self-hosted, browser-based workbench for building, curating, captioning and training LoRA and full image-model checkpoints, covering everything from raw image collection through fine-tuned model comparison in one interface.

• Models
• Generative media

A small Transformer language model and pure C inference engine for personal experimentation, running pretrained and fine-tuned models from a browser down to microcontrollers, single-board computers and routers.

• Models
• Inference

stable-pretraining is a PyTorch Lightning library that lets researchers swap and compare self-supervised learning objectives, such as SimCLR, DINO, MAE, BYOL, VICReg and LeJEPA, while keeping the same training loop, evaluation and checkpointing.

• Models

mini-verl adapts the verl reinforcement-learning framework to run PPO, GRPO and on-policy distillation training on a single consumer NVIDIA GPU, reusing existing verl configs and Hydra overrides.

• Models

A multi-agent system that automates end-to-end multimodal machine learning workflows, turning raw data into a trained ML solution with no manual pipeline building, across CLI, Python API, web UI and MCP interfaces.

• Agent frameworks
• Models

AReno is a self-contained toolkit for running reinforcement-learning post-training of language models, PPO, GRPO and SFT or DPO-style training, serving and agentic RL, on a single node without a training cluster.

• Models