Code Read the original on VentureBeat 2 min read 0

Xiaomi releases MiMo Code to challenge Claude in AI coding

Xiaomi has officially open-sourced MiMo Code V0.1.0, a terminal-native AI coding assistant designed to handle complex, multi-step software engineering tasks. The Chinese electronics giant claims the tool outperforms Anthropic's Claude Code on long-horizon benchmarks involving over 200 steps. By utilizing a specialized memory architecture and an independent checkpoint-writer subagent, MiMo Code aims to solve the common problem of AI "amnesia" during extended development sessions.

Гуманоїдний робот у чорній футболці друкує на клавіатурі в футуристичному тунелі, стіни якого вкриті моніторами з програмним кодом.
Гуманоїдний робот у чорній футболці друкує на клавіатурі в футуристичному тунелі, стіни якого вкриті моніторами з програмним кодом. · Image source: VentureBeat

According to VentureBeat, Xiaomi's MiMo AI team has released MiMo Code V0.1.0 as a terminal-native assistant that positions itself as a high-performance alternative to existing industry standards. The company asserts that the tool excels specifically in long-horizon tasks where other models often lose track of project context or previous decisions.

Advanced memory architecture and subagents

A primary differentiator for MiMo Code is its approach to "AI amnesia." While many coding agents struggle as context windows fill up, Xiaomi has implemented a cross-session memory system powered by SQLite FTS5 full-text search. This system manages four distinct layers of data: project memory via a persistent file, session checkpoints, scratch notes, and per-task progress logs.

To maintain efficiency, the system employs a unique "checkpoint-writer" subagent. This architecture functions similarly to a construction team where one agent builds the structure while a second acts as an architect, updating blueprints in real time. This allows the primary coding agent to consult structured notes when it reaches its context limits, ensuring operational momentum is not lost.

Benchmark performance and self-improvement

Xiaomi reports that MiMo Code, when paired with their MiMo-V2.5-Pro model, outperformed Claude Code in several key software engineering evaluations. The reported figures include:

  • SWE-bench Verified: 82% compared to 79% for Claude Code
  • SWE-bench Pro: 62% compared to 55% for Claude Code
  • Terminal Bench 2: 73% compared to 69% for Claude Code
  • Beyond real-time execution, the tool includes two self-improvement mechanisms. A /dream command periodically reviews and compresses historical sessions into long-term memory every seven days. Additionally, a "distill" function identifies repeated workflows from past sessions to automate them in future tasks. The project is currently available on GitHub under an MIT license for developers on macOS, Linux, and Windows platforms.

    By providing these tools as open source, Xiaomi is attempting to establish a new standard for agentic coding that prioritizes persistent state over simple context window expansion. This move signals a shift toward more autonomous, reliable AI partners capable of managing entire repositories rather than just isolated snippets of code.

    FAQ

    What is the MiMo Code checkpoint-writer subagent?
    The checkpoint-writer subagent acts like an architect in a construction team. It updates blueprints and structured notes in real time, allowing the primary coding agent to consult these notes when it reaches its context limits to maintain operational momentum.
    How does MiMo Code handle long-term memory?
    The tool includes a /dream command that periodically reviews and compresses historical sessions into long-term memory every seven days. It also features a distill function to identify and automate repeated workflows from past sessions.
    Telegram

    Fresh news on our Telegram

    Get instant alerts for new posts in «Code»

    @procodeandevenmore