Xiaomi launched MiMo Code V0.1.0 on June 10, 2026—an AI coding agent that claims to outperform Anthropic's Claude Code on a series of benchmarks. All of those performance gains come from Xiaomi's own testing, and on the free tier, every coding session sends user code to Xiaomi servers.
The tool is open-source under the MIT license and installs with a single command on macOS and Linux, or via npm on Windows. In Indonesia, MiMo Code has drawn attention only in recent days—several days after its original release—with an obvious draw: capabilities claimed to match premium subscription tools, without a dollar bill attached.
Where MiMo Code Actually Wins
The differences on short benchmarks are modest—three to seven percentage points—and all from Xiaomi's internal testing. More striking claims emerge on longer coding tasks.
On SWE-bench Verified, the most common yardstick in the coding agent community, MiMo Code scored 82% versus Claude Code's 79%—a three-point gap. SWE-bench Pro showed a wider spread: 62% to 55%. Terminal-Bench 2 put the numbers at 73% to 69%. Every figure comes from Xiaomi's own testing. TechTimes interprets the results as marketing claims rather than independent findings—neither the public SWE-bench leaderboard nor third-party testers have confirmed or disputed those scores.
A bolder claim comes from an internal double-blind A/B test: 576 developers, 474 real private repositories, 1,213 head-to-head matchups against Claude Code. Below 200 execution steps, the results ran nearly even. Above that threshold, MiMo Code's win rate climbed to 65%.
The pattern suggests Xiaomi's edge lies in context durability during extended coding sessions—the point where models typically lose the thread. MiMo Code is built on OpenCode with a four-layer memory system and SQLite FTS5-based context retrieval, engineered specifically for long sessions. Its companion model, MiMo-V2.5, claims to support a context window up to 1 million tokens.
Where Does Your Code Go?
On the free tier, code sent to the MiMo-V2.5 model moves to Xiaomi's inference infrastructure—a company operating under Chinese jurisdiction.
For individual developers with personal projects, this may not matter much. For teams at fintech firms, healthcare companies, or government agencies, the question becomes a compliance issue. Indonesia's Personal Data Protection Law and various sectoral rules constrain where sensitive data can flow—and source code carrying business logic or user data can fall within that scope.
MiMo Code works with multiple model providers, including Anthropic, OpenAI, DeepSeek, and GLM. Switching to a provider with a different jurisdiction shifts the risk point, but does not automatically answer all questions about where data goes in a user's configuration.
Here lies the irony: the segment most tempted by a free offer—startups and developers in emerging markets like Indonesia, sensitive to dollar exchange rates—is often the same segment serving clients or partners most strict about where their data lives.
DeepSeek, GLM, Now Xiaomi
MiMo Code follows the path opened by DeepSeek and GLM: releasing models or tools openly to chip away at the dominance of subscription-based products from the US. The strategy works well in cost-sensitive markets, and Xiaomi already has an organic channel to Indonesia's developer community through its established hardware footprint here.
The MiMo-V2.5 model was previously positioned as a rival to Claude Opus and Gemini. MiMo Code converts it from a model into a full harness orchestrating code reading and writing, terminal command execution, and Git repository management within a single session.
Anthropic's role as the benchmark comparison in this release comes at a fraught moment. Fable 5 and Mythos 5 were disabled days after launch following Anthropic's compliance with a US government directive, and the coding tool competition now unfolds amid geopolitical pressures that shape who can use what model, and from where. Xiaomi did not include comparisons to GitHub Copilot, Codex, or Gemini CLI in its release materials, and neither Anthropic nor Google has responded publicly to these claims.
Three signals will determine how far MiMo Code spreads in Indonesia: independent verification of the benchmark figures, clarity on data storage policies for user code once MiMo-V2.5's free period ends, and whether regulatory warnings emerge from compliance officers as the tool enters workplaces bound by data residency rules.
