Multi-Head Attention Residuals (MHAR): zero-parameter multi-head routing over the depth history — paper + reference implementation

"QUEST: Training Frontier Deep Research Agents with Fully Synthetic Tasks"

[ICML 2026 Oral] Code for "Skip a Layer or Loop It? Learning Program-of-Layers in LLMs"

Post-training with Tinker

A single CLAUDE.md file to improve Claude Code behavior, derived from Andrej Karpathy's observations on LLM coding pitfalls.

Make Any Website into CLI & Use your logged-in browser by AI agent.

A theoretical reconstruction of the Claude Mythos architecture, built from first principles using the available research literature.

[COLM '25] Single-Pass Document Scanning for Question Answering

AI handles execution, humans own the direction, and every run becomes an inspectable research artifact on disk.

Some commonly used research experiences and processes are encapsulated into Agent skills.

AI agents running research on single-GPU nanochat training automatically

This project aims to provide a high effective KV cache manage framework for llm inference and improve memory utilization and inference speed.

UniScientist is designed to advance universal scientific research intelligence through a unified paradigm

Hypernetworks that update LLMs to remember factual information

Your own personal AI assistant. Any OS. Any Platform. The lobster way. 🦞

StreamDiffusion, Live Stream APP

The simplest, fastest repository for training/finetuning medium-sized GPTs.

Lightweight Image Video Action Generation Inference Framework

implementations and experimentation on mHC by deepseek - https://arxiv.org/abs/2512.24880

Official JAX implementation of End-to-End Test-Time Training for Long Context

TurboDiffusion: 100–200× Acceleration for Video Diffusion Models

[ECCV 2026 Oral] Implementation of "Live Avatar: Streaming Real-time Audio-Driven Avatar Generation with Infinite Length"

A PyTorch-native inference engine with cache, parallelism, quantization and cpu offload for DiTs.

Official codebase for "Self Forcing: Bridging Training and Inference in Autoregressive Video Diffusion" (NeurIPS 2025 Spotlight)

Accelerating MoE with IO and Tile-aware Optimizations

A compact implementation of SGLang, designed to demystify the complexities of modern LLM serving systems.

A Reproduction of GDM's Nested Learning Paper