MiniMax Launches Code 2.0 to Slash Latency by 90%
MiniMax has released Code 2.0, rebuilding its desktop coding agent on the open-source Pi Agent framework to slash latency by over 90% and introduce flexible model integration.

MiniMax has launched Code 2.0, a major overhaul of its desktop coding agent for macOS and Windows. By rebuilding the application's core architecture on the open-source Pi Agent framework, the company has reduced p95 and p99 first-token latency by more than 90 percent. This transition aims to improve long-task stability and responsiveness, addressing common performance bottlenecks in developer workflows.
The underlying Pi Agent framework, created by Mario Zechner, is a minimalist open-source project with over 46,000 GitHub stars. It relies on a TypeScript SDK and just four core tools: read, write, edit, and bash. By adopting this lean foundation, MiniMax avoids the overhead of complex commercial tooling, prioritizing a controllable agent loop and a short system prompt to maximize reliability.
Beyond the performance gains, Code 2.0 introduces several user-facing features designed to streamline development. A new Remote Control feature allows developers to monitor agent progress, send instructions, and approve permissions directly from their mobile phones. The update also integrates a built-in browser in the sidebar, enabling inline previews and editing of HTML output without switching applications. To accommodate different users, the interface now splits into a full developer context mode for engineers and a simplified progress view for non-technical team members.
For practitioners, the update offers significant flexibility through Bring Your Own Key support. Users can connect to any AI provider, including Anthropic, OpenAI, Gemini, or locally hosted models, using a base URL and API key. This allows developers to leverage their preferred models while benefiting from the agent's low-latency desktop environment and mobile monitoring capabilities.
This is our own summary of reporting by AlphaSignal



