Colibri is a pure-C, zero-dependency inference engine that runs frontier Mixture-of-Experts models like GLM-5.2 (744B parameters) on consumer hardware with only about 25 GB of RAM by streaming routed experts from NVMe disk on demand. It exploits the fact that only ~40B parameters activate per token and only ~11 GB of experts change between tokens, treating VRAM, RAM, and SSD as one staged hierarchy with a per-layer LRU cache and optional pinned hot-store. The engine implements Multi-head Latent Attention with a 57x-compressed KV-cache, AVX2-optimized int8/int4 dot-product kernels at 119 GFLOP/s, and multi-token speculative decoding. Released under Apache 2.0, Colibri lets researchers and builders hold a frontier-class model on their own machine, probe and measure every expert firing in real time, and contribute optimizations. It matters now because it collapses the gap between datacenter-scale MoE inference and a mid-range PC.
Reader rating
No ratings yet
You might also like
codemap is an MIT-licensed project brain for AI coding tools that gives LLMs instant architectural context from your codebase without burning tokens. It generates a fast tree/context view, dependency flow, dependency blast-radius analysis, and a layered handoff format for cross-agent continuation, then exposes everything through a JSON context bundle and an MCP server compatible with Claude Code and Codex. A built-in Codex plugin and community skill registry make it easy to install and share. Developers use codemap to onboard agents to large repos in seconds, keep session continuity across handoffs, and scope the impact of a change before running it.
Ollama is a local AI platform for running, managing, and sharing open models on your own machine or private infrastructure. It makes it easy to pull models, serve them through an API, and integrate local inference into developer workflows without relying on a fully managed cloud stack. Teams use Ollama for privacy-sensitive assistants, internal tools, offline experimentation, and rapid testing of open-weight models across laptops, workstations, and servers. It is especially useful for developers, operators, and AI builders who want quick setup with less operational overhead. What makes Ollama distinctive is how approachable it is: it packages model runtime, distribution, and deployment into a streamlined experience that helps people get productive with local AI in minutes instead of spending days on configuration.
FileForge Finder is an AI-powered local file search utility that optimizes search results for developer workflows. It uses natural language processing to understand query intent and prioritize relevant files, code snippets, and documentation. The tool integrates with popular IDEs and terminals to provide instant, context-aware file retrieval, reducing time spent navigating complex project structures. It supports multiple file formats and offers advanced filtering by content type, modification date, and relevance.
From the blog
NVIDIA's $12B Poolside deal reveals a new M&A structure: IP licensing + talent transfer + founder retention. Compute scarcity is rewriting the playbook for frontier AI labs…
Amazon is destroying rare books for AI training data. Robin Williams' family revived his Instagram to fight AI likeness abuse. These are the same story — AI treating culture as free raw material.
ElevenLabs in Claude and Cursor Origin show why the next useful AI interface may be the one that safely runs the workflow…
