The ecosystem of AI coding assistants has a new contender. Xiaomi has released MiMo Code V0.1.0, a terminal-native tool that claims to outperform Anthropic's Claude Code on complex tasks involving over 200 steps. The innovation lies not just in benchmark numbers but in the architecture: a persistent memory system that tackles the context window degradation plaguing long coding sessions.
Structured memory for AI agents
Anyone who has used an AI agent for programming knows the issue. After dozens of steps, the model loses track of earlier decisions, conventions, and project state. Xiaomi built MiMo Code to overcome this with a mechanism inspired by SQLite FTS5 full-text search. The system writes checkpoints, notes, and progress logs into a four-layer persistent memory: project memory (a MEMORY.md file), session checkpoints, scratch notes, and per-task logs. An independent subagent, the checkpoint writer, constantly updates this data without interrupting the main agent. When the context window approaches its limit, the main agent consults the subagent and rebuilds the environment, preserving operational momentum.
Sponsored Protocol
Benchmarks: dominance on long tasks
The figures released by Xiaomi are striking. Paired with the MiMo-V2.5-Pro model, MiMo Code scored 82% on SWE-bench Verified (vs. Claude Code's 79% with Claude Sonnet 4.6), 62% on SWE-bench Pro (vs. 55%), and 73% on Terminal Bench 2 (vs. 69%). The difference is not solely model-based: running the same MiMo-V2.5-Pro in both harnesses, MiMo Code outperforms Claude Code by roughly five percentage points on each benchmark, proving that agent architecture matters as much as the underlying model. In a human double-blind A/B evaluation involving 576 developers, the two systems split roughly 50/50 under 200 steps, but past 200 steps MiMo Code wins over 65% of the time. These are vendor self-reported numbers not yet independently verified, but the trend is clear.
Sponsored Protocol
Efficient models and aggressive pricing
MiMo Code integrates with the MiMo-V2.5 model, released in April 2026. With 310 billion total parameters (15 billion active per inference) and a 1-million-token context window, the model costs $0.40 per million input tokens and $2.00 per million output tokens. The Pro version, with 1.02 trillion parameters, runs $1.00/$3.00 up to 256K context, scaling to $2.00/$6.00 beyond. With cache hits, input costs drop to as low as $0.20. Xiaomi also bundles free limited-time access to MiMo Auto, requiring no registration. For developers who prefer other models, MiMo Code supports third-party backends including DeepSeek, Kimi, and GLM, plus any OpenAI-compatible API.
Sponsored Protocol
Voice, compose mode, and compatibility
MiMo Code includes features designed for real developer workflows. With Compose mode, pressing Tab activates an autonomous agent that designs, plans, writes code, tests, and reviews, following a heavy-planning-upfront, stable-verification-later strategy. Voice control, based on MiMo-ASR and TenVAD, allows hands-free dictation and command execution. The tool automatically imports MCP servers, custom skills, and API configurations from Claude Code, making migration painless.
Implications for enterprises and developers
For engineering teams, MiMo Code is a low-risk, high-value evaluation candidate: MIT license, inspectable architecture (fork of OpenCode), and bring-your-own-model support. The persistent memory system addresses a real pain point. Caveats include the limited-time free access, data routing through Xiaomi servers (critical for data-residency policies), and the V0.1.0 release signaling immaturity. Teams subject to U.S. government restrictions on Chinese vendors should weigh the geopolitical context. To explore more about AI agents, read our complete guide on AI Agents and Advanced Automation. For another example of open-source AI tooling, check out Stable Diffusion Locally, showing how local customization is transforming development.
Sponsored Protocol
The terminal AI coding agent war is just beginning, and Xiaomi is playing the open-source, low-cost card. The next step will be whether the developer community validates the claimed numbers. For more background on Xiaomi, visit the Wikipedia page.
