Language Models

NInfer is a from-scratch C++/CUDA engine that trades all generality for maximum single-GPU throughput on a closed registry of Qwen checkpoints.

VoiceMem exists because real-time voice agents need to remember who you are, what you said, and how you felt—without making you wait.

This repo is a surgical stack of vLLM patches, requantization scripts, and speculative decoders built to squeeze Qwen3.8-27B — and up to 268k tokens of context — into a single 24 GB consumer GPU.

It exists to replace a tangle of single-modal encoders with one model that puts text, images, videos, and documents into the same vector space.

Because serving DeepSeek V4 Flash at 1M-token context across two DGX Spark nodes takes more than a `docker run`.

This tool launders AI-generated text through a chain of translations and LLM rewrites to throw detector tools off the scent.

It turns a Raspberry Pi 5 into a fully offline kiosk for two people to talk across languages.

It exists because Persian TTS and linguistics tools need a modern, learned grapheme-to-phoneme converter that outputs standard IPA without brittle hand-written rules.

Soup exists because fine-tuning LLMs shouldn't require a cloud budget, SSH, or a PhD in distributed systems.

It unifies Stable Diffusion, GGUF chat, Whisper, and Kokoro TTS into a single offline desktop GUI so you can skip cloud APIs, subscriptions, and censorship filters.

Xime is a deliberately minimal, Rime-based Android input method that serves as its author's personal testbed for on-device AI experiments in predictive text and speech recognition.

RedKnot accelerates long-context inference by sorting attention heads into four species—global, local, retrieval, and dense—then giving each its own KV reuse strategy and sparsity rules, built as an SGLang extension.

It breaks complex tasks across a team of specialized LLM agents that refine their own skills as they work.

A hands-on Node.js tutorial series that makes you implement embeddings, vector stores, and retrieval yourself so RAG stops feeling like magic.

MTPLX squeezes extra tokens per second out of Apple Silicon by using the multi-token prediction heads that ship with modern models like Qwen 3.6, instead of leaving them idle like most runtimes.

Twenty-five bite-sized projects showing how to wire up LLMs, RAG, and agents into things that actually do work.

ChatGPT2API exists to reverse-engineer the ChatGPT web interface into a self-hosted OpenAI-compatible API, complete with disposable account pools and a Vue management console.

Squeezes datacenter-scale MoE inference onto consumer GPUs by treating your desktop’s heterogeneous resources as a unified, elastic runtime.

To prove that a 2.78-trillion-parameter model can run on a single CPU with 8 GB of RAM and no GPU.

It gives open-source LLMs a secure, sandboxed Linux desktop they can click, type, and shell into while you watch.