Beyond Claude Code: MiniMax CLI’s Bizarre Evolution Breaks All AI Rules

When a company releases a powerful AI tool and then changes the rules after the fact, things can get messy. That’s exactly what happened with MiniMax and its new CLI tool, MMX-CLI. The tool launched on April 9, 2026, and it’s already stirring up controversy.

MMX-CLI is a command-line tool built for AI agents. It works with environments like Claude Code and OpenClaw. It gives users access to MiniMax’s multimodal AI features through simple command execution. The tool was designed to handle complex automated workflows without getting in the way of how agents read and process data.

The tool has some smart technical features. It sends progress updates to one channel and clean data to another. This keeps outputs organized and easy for machines to read. It also uses specific error codes so agents know exactly what went wrong. A special flag lets long tasks run without blocking other processes.

Then came the M2.7 model. MiniMax open-sourced it on April 12, 2026, just days after the CLI launch. It’s a massive 230-billion-parameter model that uses a Mixture-of-Experts design. What makes it unusual is that it helped develop itself. It can analyze its own failures, update its memory, and build new skills across training rounds. It ran over 100 rounds of self-improvement focused on programming tasks.

The numbers are impressive. M2.7 scored 1495 on the GDPval-AA benchmark, the highest among open-source models. It followed complex skill instructions 97% of the time, even when those instructions were longer than 2,000 tokens.

But then MiniMax updated the license. After release, they required commercial users to get authorization before using the model. That change didn’t go over well. Hundreds of critical comments flooded the Hugging Face discussion page. The company said the update was meant to stop hosting providers from distributing degraded versions of the model.

The license still allows internal fine-tuning and research without approval. But it’s stricter than open licenses like MIT or Apache. It’s a pattern that’s become common among Chinese AI labs. Despite the licensing controversy, M2.7 matches benchmark scores of GPT-5.3-Codex on SWE-bench Pro, where it achieved a score of 56.22% across complex software engineering problems. MMX-CLI was built on MetaEra, further tying the tool’s infrastructure to a platform designed to support AI agent environments at scale. The researchers behind M2.7 relied heavily on machine learning algorithms to drive the model’s self-improvement cycles across hundreds of training rounds.