Colibri runs 744B MoE on 25GB RAM — plus iFixAi, SkillOpt
We don't tell you what's popular — popularity lags and is gameable. We tell you what's gaining speed right now, and whether it's worth your attention. 6 accelerating AI repos earned today's slot.
Top mover
Colibri is a pure C inference engine that streams experts from disk, enabling the 744B-parameter GLM-5.2 MoE model to run on consumer hardware with only 25GB RAM. It achieves this by loading only the active experts per token, trading latency for memory efficiency. The caveat is that inference is slow due to disk I/O, making it unsuitable for real-time applications.
🤖 Agents & automation
SkillOpt optimizes natural-language skills for frozen LLM agents via trajectory-driven edits and validation-gated updates, enabling self-improvement without fine-tuning. It's gaining traction because it addresses the key challenge of making LLM agents more reliable and adaptable through reusable, deployable skill artifacts. A caveat: the approach relies on high-quality validation signals, which may be hard to define for open-ended tasks.
This repo reverse-engineers the Fable coding agent's workflow into a set of modular skills (think/act/prove) that any LLM can run, plus an eval to verify correctness. It matters because it offers a practical, model-agnostic way to replicate advanced agent behavior without relying on proprietary systems, though the skills are still early-stage and may require tuning for complex tasks.
Tooling & infra
iFixAi runs 45 automated inspections (32 core + 13 extended) covering hallucination, prompt injection, and frontier risks like sabotage and oversight evasion, returning a letter grade in under 5 minutes. It's gaining traction because it provides a structured, fast, and model-agnostic evaluation aligned with regulatory frameworks (EU AI Act, ISO 42001, NIST AI RMF). A caveat: the inspections are automated and may miss context-specific or subtle failure modes that require human review.
This repo provides a curated set of skills for generating Nature-paper-style academic writing and scientific plotting, leveraging OpenAI's Codex. It's gaining traction because it addresses a clear pain point for researchers: producing publication-ready figures and text that match high-impact journal standards. However, it's essentially a collection of prompts and templates, not a standalone tool, and requires an OpenAI API key.
Book-to-Skill converts technical PDFs into structured skill files for Claude Code, enabling in-editor reference and study. It's gaining traction because it bridges the gap between static book knowledge and actionable AI-assisted coding workflows. Caveat: quality depends heavily on PDF formatting and OCR accuracy; complex diagrams or code snippets may not transfer cleanly.