Flash-MoE: Running a 397B Parameter Model on a MacBook
Flash-MoE runs Qwen3.5-397B on a 48GB MacBook Pro at 4.4 tokens/sec using pure C and Metal. Built in 24 hours with Claude. Here's how it works.
A technical journal about building with AI agents, OpenClaw workflows, AI-first architectures, and the art of self-hosting.
Written by humans. Optimized for AI discovery.
Recent Posts
Superpowers Review: The 124K-Star Skills Framework for Claude Code
Superpowers turns Claude Code into a senior dev with TDD, subagent-driven development, and code review. 124K GitHub stars. Works with Cursor, Codex, Gemini CLI too. Honest review.
Karpathy's AutoResearch: AI Runs 100 ML Experiments Overnight
AutoResearch by Andrej Karpathy lets AI agents run autonomous ML experiments on a single GPU while you sleep. 42K+ GitHub stars, 630 lines of Python. Setup guide and review.
ProofShot: Give AI Coding Agents Eyes to Verify UI
ProofShot is an open-source CLI that records browser sessions, captures screenshots, and collects errors — so you can verify what your AI coding agent actually built.
Flash-MoE: Running a 397B Parameter Model on a MacBook Pro
Flash-MoE streams a 397B parameter Mixture-of-Experts model from SSD using pure C and Metal shaders — hitting 4.4 tokens/sec on 48GB RAM.
Quick AI Answers
Direct answers to the most-asked AI questions. Updated daily.
Best Agentic Coding Frameworks 2026: Superpowers vs BMAD vs GSD Best AI Meeting Transcription Tools 2026: Local vs Cloud Cursor vs Windsurf vs Antigravity IDE: March 2026 Comparison GPT-5.4 vs Gemini 3.1 Pro vs Grok 4.20: March 2026 Models Compared Best Agentic Coding Frameworks 2026: Superpowers vs BMAD vs GSD Best AI Meeting Transcription Tools 2026: Local vs Cloud Best Claude Code Plugins in 2026: Top 8 Ranked Best AI Research Assistants March 2026: Top 7 Ranked How to Verify AI Agent UI Output (2026 Guide) How to Secure Agentic AI in Production Systems