Drex Max v1: an open-weight 27B decision model you run locally
It returns probabilities over 255 choices instead of text, fits a 24GB Mac, free under $1M revenue — and Claude Code reportedly stops cloning games.
Copy markdown
A decision model you run, not a chatbot you prompt
Nace AI open-sourced Drex Max v1, a dense 27B model that reads a state — text, JSON, or images — plus your typed questions and returns a probability for each of up to 255 options in one forward pass; nothing is generated. The 4-bit build runs on a 24GB Apple Silicon Mac via llama.cpp, Ollama, or MLX (the 8-bit GGUF wants 48GB) with a 262K-token context, so ticket routing, tool selection, and agent guardrails can run locally without parsing model prose or paying per token.
The license has a $1M ceiling
Drex Max ships under Open-RAIL-M, not a standard open-source license: free commercial use for individuals and orgs under $1M in trailing-12-month revenue or $1M raised, with personal and research use exempt. Cross either threshold — or build products that compete with Nace's decision or document-intelligence tools — and you need a written deal. Check it before wiring the weights into anything you charge for.
Claude Code reportedly stops cloning games
Developers report an internal flag added two lines to Claude Opus 5.5's Claude Code system prompt telling it to refuse recreating existing games, with Addy Osmani confirming the change on X. Some say the guardrail also catches legitimate reverse-engineering work. If your agent suddenly balks at a game-adjacent build, this is the likely cause.