LoMa: Local Feature Matching Revisited

Efficient Universal Perception Encoder: a single on-device vision encoder with versatile representations that match or exceed specialized experts across multiple task domains.

Embedding Vector Oriented Clustering

a disk cache for using DuckDB to access Data Lakes (ducklake, iceberg, delta)

⚡ Super fast clustering for high-dimensional vectors on CPUs (x86, ARM) and GPUs — for Python and C++. 100x faster clustering of vector embeddings than FAISS

MDM-Prime-v2: Binary Encoding and Index Shuffling Enable Compute-optimal Scaling of Diffusion Language Models

Fast and memory-efficient exact kmeans

Build ultra fast, tiny, and cross-platform desktop apps with Typescript.

Pure MLX implementations of UMAP, t-SNE, PaCMAP, TriMap, DREAMS, CNE, MMAE, and NNDescent for Apple Silicon. Metal GPU for computation and video rendering.

A Quirky Assortment of CuTe Kernels

UMAP in pure MLX for Apple Silicon. 30x faster than umap-learn.

100M tokens. Infinite compute. Lowest val loss wins.

Official implementation of VLANeXt.

The Adana algorithm official repository

Algorithms for latent compaction

Various ML tidbits in Python/PyTorch and C++

Official implementation of ViT-5: Vision Transformers for The Mid-2020s

Code and models for the paper: Hybrid Linear Attention Done Right: Efficient Distillation and Effective Architectures for Extremely Long Contexts

Helpful kernel tutorials and examples for tile-based GPU programming

FLA but cuTile

LLaDA2.0 is the diffusion language model series developed by InclusionAI team, Ant Group.

nanoTabPFN in X minutes

An interface library for RL post training with environments.

Kernels, of the mega variety :)

Qwen3-0.6B megakernel: 527 tok/s decode on RTX 3090 (3.8x faster than PyTorch)