MiniMax open-sources MiniMax Code, an MIT terminal coding agent

A Claude Code-style TUI you can fork and point at any OpenAI- or Anthropic-compatible model — including MiniMax's own open M2.1 at a tenth of Sonnet's price.

Nowline SEP 22 2:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • An open coding agent you can fork

    MiniMax Code is a terminal agent that reads your repo, edits files, and runs tests from a TUI with permission prompts. First-party code is MIT-licensed and live on GitHub at ~1.7k stars on day one; you install it with a single curl-to-bash line.

  • Bring your own model — it's not locked to MiniMax

    Point it at a MiniMax account, or at any OpenAI- or Anthropic-compatible endpoint, including self-hosted and third-party models. That makes it a genuinely hackable alternative to Claude Code, Codex, and Gemini CLI rather than a captive client.

  • Headless mode and ACP, for CI and editors

    An `mcode exec` mode runs non-interactively for scripts and CI/CD pipelines, and Agent Client Protocol support wires the agent into compatible editors. Build this weekend: a pipeline step that opens a PR fixing your failing tests.

  • The model behind it: M2.1, open and cheap

    MiniMax's open-weight M2.1 (Modified-MIT) runs $0.30/$1.20 per million tokens with a 205K context. It scores 74.0 on SWE-bench Verified and 47.9 on Terminal-Bench 2.0 — within reach of Claude Sonnet 4.5 (77.2 / 50.0) at roughly a tenth of the price.

  • The catch: the desktop app stays closed

    The open repo covers the TUI, headless CLI, and ACP layers only — the MiniMax Code desktop app ships separately and isn't included. Bundled files also keep their own upstream licenses under the MIT wrapper, so read the notices before shipping.

  • Elsewhere: a 100M image model you can train in a day

    SupraLabs open-released Supra2, a 100M-parameter text-to-image model reportedly trained in under 10 hours on a single H100. Weights are on Hugging Face and already in the LocalAI gallery — small enough to fine-tune on your own GPU.