low-memory

Tiny engine, immense models — run large MoE LLMs (gpt-oss, Mixtral, Qwen3-MoE) on ordinary machines by streaming experts from disk. OpenAI-compatible server with tool calling + hybrid cloud relay; CPU, Apple Silicon (MLX) & CUDA.

(consider giving us a star!) Lightweight, feature‑packed DeepSeek app built with PyWebView, no Electron bloat. Includes all features of the most‑starred builds plus custom UI polish, dynamic greetings, animations, auto‑updater, enhanced Markdown rendering, and a injection js with lots of features :] <3

Run MiniMax H3 on 16 GB Apple Silicon Macs with layer-by-layer MLX weight streaming.

A lightweight and memory-efficient HTTP client module for MicroPython, optimized for use cases on resource-constrained devices such as ESP32.

Pingstream is a lightweight, Streamlit-based API testing tool using curl under the hood. Designed for low-RAM systems, it supports GET, POST, PUT, DELETE with JSON, headers, file uploads, and Postman/OpenAPI import.

Fine-tune frontier MoE models on hardware you already own — experts streamed from disk, LoRA on the hot path. 🐣

Zero-memory, multi-core-CPU local AI training for ordinary laptops.

Prompt-boundary disk relay, review, resume, and assembly for memory-constrained MiniMax H3 long-video chains in ComfyUI

Surgical reasoning on consumer silicon. Hybrid SSM + causal memory architecture with entropy-gated System 1/2 dispatch, O(1) inference memory, and continual learning — designed for 16 GB VRAM.

Efficient Pairwise Cosine Similarity Computation

Run an agentic coding LLM on the GPU you already own. No cloud, no invoice — just llama.cpp on Vulkan, a few glue scripts, and a 7–8B model that calls tools. This repo turns the whole thing into a LAN OpenAI-compatible endpoint that opencode (or any OpenAI client) can drive.

Low-memory reproduction of DORA test-time scaling on a single 8GB GPU

Memory-safe sequential long-video upscaling with audio preservation for ComfyUI / 長尺動画を逐次処理・音声維持で高解像度化

Homeassistant light chromium browser

Run MiniMax Music 3 locally on Apple Silicon Macs with MLX, phase residency, and an Idea-to-Music workflow.

极轻量的本地 mp3 播放器 · Windows MCI 原生播放 · 空闲内存约 37 MB，打游戏后台听歌不卡顿 · 零第三方依赖

Eidetic — a single-process, single-database memory server for AI agents. SQLite-native: vectors + FTS5 + knowledge graph + raw logs. Five-layer memory (L0 raw → L1–L4), three-tier recall, auto-healing. One-command deploy, local-first, works with any MCP-compatible agent (OpenClaw, Claude Code, etc.).

A minimal Python shell optimized for low memory usage (~21 MB RAM). Features tab completion, command history, I/O redirections, and built-in commands compiled to a native binary with Nuitka.