Explore the directory

A comprehensive Chinese-language tutorial on deploying large language models locally, covering how transformer inference and quantization work, choosing hardware, and comparing deployment tools such as LM Studio, Ollama, llama.cpp, vLLM and TensorRT-LLM.

• Tutorials

A fork of llama.cpp that adds TurboQuant, custom low-bit quantization formats with CUDA kernels for smaller memory use, and TriAttention, GPU-accelerated KV-cache pruning that scores cached tokens by importance to keep long-context inference inside a fixed memory budget.

• Inference

A learnable memory framework for LLM-powered multi-agent systems, accepted to EMNLP 2026, that trains a memory composer to distill an experience bank of raw interaction trajectories into compact, role-aware latent memories for each agent.

• Agent memory
• Papers

An educational project that trains a large language model from scratch through two pipelines: a native C++ engine for data ingestion, tokenization and training, and a Python pipeline using Apple's MLX framework that scales a decoder-only GPT model to 1.6 billion parameters.

• Models
• Tutorials

A durable AI coding agent for the terminal, built on an embeddable TypeScript runtime kernel capped at 2,200 lines, that resumes exactly where it left off after a crash because every approval and tool result is committed to disk as it happens.

• Coding agents

A fully local, multi-model AI agent framework that routes tasks such as planning, coding, reasoning and vision to specialized models running under Ollama, then synthesizes their answers through a Mixture-of-Agents aggregator, tuned for consumer GPUs such as the RTX 4080 or 5080.

• Agent frameworks
• Inference

A hands-on workshop repository for building, testing and deploying single-agent and multi-agent Hosted Agents to Microsoft Foundry Agent Service, entirely from Visual Studio Code, using the Microsoft Foundry extension and the Python Agent Framework.

• Tutorials

A shared, content-addressed and cryptographically signed external memory store that lets independent AI agents cite and verify the same real-world observation, such as a satellite image or a camera record, without sharing a model or a vendor.

• Agent memory

A collection of GitHub Copilot customizations, including specialized agents, coding-standard instructions, prompt templates and Model Context Protocol servers, built for the Norwegian government's own technology stack.

• MCP servers
• Agent skills

A curated list of AI agent frameworks, platforms, infrastructure, research papers and learning resources, organized by frameworks and libraries, platforms, evaluation and safety, tutorials, and use cases.

• Curated lists

A retrieval-augmented generation framework, first built at MIT for the SubMIT computing project, aimed at research and education teams who need a configurable, private and extensible assistant with a low barrier to entry.

• RAG

A vector similarity search engine written in Rust that compiles to a 54 KB WebAssembly module, letting a web application run semantic search entirely in the browser with sub-millisecond query latency and no server, API key or per-query cost.

• RAG
• Databases

A cross-platform CLI written in Rust that keeps one set of AI coding-agent instructions, commands and skills in a single directory and symlinks them out to the different locations that tools such as Claude Code, GitHub Copilot, Cursor and Codex each expect.

• Agent skills
• CLI tools

A multi-agent team designer that interviews the user, decomposes the work into responsibilities, and generates a ready-to-run set of agents, skills and rules for either Claude Code or Codex, with configuration formats convertible between the two.

• Agent frameworks
• Agent skills

A local chat application and HTTP gateway for running large language models entirely on a personal machine, with a model catalog for verified downloads and on-demand vision support for images and PDFs.

• Inference

A custom wake-word model for Home Assistant voice assistants, built with microWakeWord, that runs entirely on-device and comes with its full training pipeline so a user can train the same setup on any name of their own.

• Speech

A benchmark for evaluating multimodal large language models on front-end code generation, covering React, Vue, Angular and plain HTML and CSS across generation, editing and bug-repair tasks on nine hundred real webpage samples.

• AI evals

A human-in-the-loop workbench that turns a domain expert's intuitive sense of "find me content like this" into a reproducible, auditable classifier trained on a rubric-judged sample of a large text corpus.

• RAG
• Data engineering

A memory-optimization toolkit for fine-tuning SDXL-class and ANIMA diffusion models on 12 GB consumer GPUs, combining a custom optimizer that reduces optimizer-state memory with a per-layer scheduler that skips computation on layers whose gradients have gone flat.

• Models
• Generative media

A unified vLLM serving configuration, distributed as a Docker Compose stack, for running large language models on NVIDIA DGX Spark GB10 hardware, supporting either a single-node setup or two nodes joined over a 200 Gbps RDMA link.

• Inference
• Containers

An unofficial PyTorch implementation of the Recurrent Looped Transformer proposed by Yifan Zhang of Princeton, an encoder-decoder architecture that can share weights between encoder and decoder and vary its recurrent block size across a sequence.

• Models
• Libraries

A reinforcement-learning framework for language models that builds adaptive training environments where a model practices problem-solving and is scored with multi-dimensional, partial-credit rewards instead of a single binary right-or-wrong signal.

• Models

A framework for fine-tuning GPT-2 to generate math riddles and their solutions in multiple explanatory styles, using a dual-decoder architecture and a coherence-penalty loss that ties a generated riddle's wording to its generated answer.

• Models

An MCP server that lets Claude or another MCP client operate a running OrcaSlicer instance: loading models, changing settings, slicing a plate and reading back numeric, per-feature print-time and material breakdowns instead of only a picture of a legend.

• MCP servers