MiMo-Code na Xiaomi alternative to Claude Code?

Wetin be MiMo-Code?

MiMo-Code na terminal coding agent wey Xiaomi MiMo team build and release as open source for June 2026 under MIT license. Xiaomi launch write-up wey dem date 10 June 2026 don already describe am as open-sourced. E dey run as CLI (command line interface) tool for your shell. E dey read and edit files for project, run shell commands, control git, and keep notes about the codebase between sessions.

The structure na the same as Claude Code. You go describe task with plain language, the agent go suggest edits and commands, and you go approve dem. Two differences dey important. The source dey open, so you fit read exactly wetin the agent send go the model. And the model na separate choice from the agent, so the same harness fit connect to Xiaomi models, OpenAI-compatible endpoint, or model wey you host by yourself.

The repository name na MiMo-Code and the command wey e install na mimo. Both names dey for the documentation, so if you search for any one of dem, you go find the same tool.

MiMo-Code come from where, and why e matter

MiMo-Code na fork of OpenCode, wey be the open-source terminal agent from the sst project. The LICENSE file still get the original Copyright (c) 2025 opencode line, and the README get short section about the relationship. One open issue dey for the repository, asking Xiaomi to give the OpenCode contributors fuller credit than wetin one preserved license line fit give.

When you know this, one thing wey otherwise go look impossible go make sense. Project wey first release for June 2026 already ship polished TUI (text user interface), plugin system, provider abstraction wey fit talk to many model APIs, MCP (model context protocol) support, and LSP (language server protocol) integration. Dem no do all this work within few weeks. E come with the fork. The new things from Xiaomi na the memory layer and the connection to Xiaomi own models.

If you don already run OpenCode for VPS, the layout and keybindings go feel familiar, because na the same code dey underneath.

Wetin "models and agents co-evolve" really mean

The repository tagline be "Where Models and Agents Co-Evolve". That phrase get two separate claims, but na only one get published detail to back am.

The first claim concern runtime, and dem don document am. MiMo-Code dey keep persistent memory store for each project, backed by SQLite full-text search, together with session state and task progress. Two slash commands dey drive am. /dream dey pull durable knowledge from recent sessions and write am into project memory. /distill dey turn procedure wey dem repeat into reusable skill. Xiaomi own engineering write-up talk say Dream dey trigger automatically every 7 days and Distill every 30 days, instead of after every session. So the agent behaviour for your repository dey change over time while the model file remain the same. The store dey fill by itself, but e no replace instruction files wey you write by hand. For monorepo, e better make dem be one small file for each package instead of one root file wey the agent go load based on the directory e dey work in. Writing those files na the easy half. When the agent pass one rule wey you sure say clear, the reasons why instruction fit get ignored dey more mechanical than mysterious. Memory distilled from past sessions na only one way to stop agent from relearning project from zero. A tree-sitter map of the repository wey agent fit query through MCP dey attack the same waste from the other side, by giving am the structure instead of making am read through everything again.

The second claim concern training, and for here you need dey careful. Xiaomi describes the post-training of MiMo-V2.5 as a three-stage pipeline: supervised fine-tuning, then large-scale agentic reinforcement learning (RL), then on-policy distillation from several specialist teacher models into one student. Agentic RL mean say dem train the model with multi-step tool-calling episodes instead of single question-and-answer pairs. That one real, and na why the model fit hold together across long tool sequences. Wetin Xiaomi never publish na training report wey show say dem optimise the model inside this specific harness, with the harness changing as response.

So understand "co-evolve" as statement of intent from team wey own both sides. The benefit wey you fit actually observe dey narrower, but e still get value: the same people write the tool schemas and system prompt and ship the model, so dem fit each other. Claude Code no fit work like that, because its harness na client of fixed API model. Anthropic dey train its models for agentic use too, but you no fit swap the model from underneath the harness.

MiMo-Code session dey look like this for terminal

When you run mimo, e go open full-screen terminal interface. You go type task, agent go read files, propose change, then ask before e write or run anything. Apart from default build mode wey e ship with, e get plan mode and compose mode, plus subagent system wey fit hand part of job to separate agent run. Claude Code reach the same level from the other side. There, one output style dey rewrite the system prompt itself instead of switching you between fixed set of prepared modes. E also get plenty built-in skills. Dem be prepackaged procedures wey agent fit call by name. Skill fit also change how agent dey work instead of wetin e know. Na this idea dey behind Ponytail, one skill wey make agent stick to the smallest change wey go work.

You need know some details before you start.

• By default, approval dey happen for each action. Flag --dangerously-skip-permissions remove the prompts. The name correct, so use am only where mistake no matter. MiMo-Code make am all-or-nothing switch, while Claude Code divide the same decision into permission modes with different levels. So, before you use the flag, wetin auto mode allow agent do without asking na the useful comparison.
• E fit run without interactive terminal. mimo run dey take prompt and execute non-interactively. Na this script or CI job need.
• Configuration dey inside files wey you fit commit. Project dey read .mimocode/mimocode.jsonc, while global settings dey inside ~/.config/mimocode/mimocode.jsonc.
• Credentials no dey inside those files. Dem dey inside ~/.local/share/mimocode/auth.json, so keep that path out of git and out of any backup wey you share.

Agent wey fit run shell commands na program wey you don give control of your shell. Give am its own machine and its own user. The safest way to test new one na disposable VM wey you fit destroy after the session, because bad command go cost you rebuild instead of restore.

MiMo-Code really dey beat Claude Code?

Xiaomi launch material make one specific claim: run the same model for both harnesses, and MiMo-Code score higher. These na the figures wey dem publish.

[
  {
    "tool": "MiMo-Code",
    "swe_bench_pro_pct": 62,
    "terminal_bench_2_pct": 73
  },
  {
    "tool": "Claude Code",
    "swe_bench_pro_pct": 57,
    "terminal_bench_2_pct": 68
  }
]

For SWE-bench Pro, Xiaomi report 62 percent for MiMo-Code against 57 percent for Claude Code. For Terminal Bench 2, the figures na 73 percent and 68 percent. The framing be say dem keep the model constant, so gap of about five points come from the agent system, no be from the model. This framing only get meaning if you know wetin agent harness actually cover: the loop, the tool definitions, the permission prompts, and the session state na the parts wey differ between the two, while the weights remain the same.

Treat those ones as vendor numbers. Xiaomi run the evaluation and publish am, and as of August 2026 MiMo-Code no dey appear for the public SWE-bench or Terminal-Bench leaderboards. So for that time, nobody independent don reproduce the gap. Xiaomi dey unusually direct about the limit of its own result. The company write-up talk say these benchmarks measure one-shot problem solving for single repository issues, and say for its own testing, the two agents perform about the same below roughly 200 execution steps. The advantage wey dem claim dey show for long runs, and one-shot benchmark no cover that case.

Na this be the honest reading. Five-point gap for benchmark wey vendor run na reason to try the tool. E no be reason to move the work wey your team depend on.

MiMo-Code cost wey e go take run

When Xiaomi launch am, e ship one free channel wey dem call MiMo Auto, and person fit use am without registration. From the beginning, dem describe am as limited-time offer. On 23 July 2026, Xiaomi announce say MiMo Code free-use phase go end by 18:00 on 26 July 2026, and dem direct users to paid Token Plan instead. That channel don close, so plan for paid access.

Paid access to Xiaomi platform cheap compared with frontier providers. One permanent price cut wey synchronize globally take effect on 27 May 2026. E set MiMo-V2.5-Pro at $0.435 per million input tokens when cache miss happen, and $0.87 per million output tokens. The same cut remove separate pricing based on input length, so one rate cover the full 1M token context window. MiMo-V2.5-Pro don retire since then. Xiaomi route requests for am to MiMo-V2.6 from 18:00 on 14 October 2026 (UTC+8), and dem retire MiMo-V2.5-Pro and MiMo-V2.5 at 10:00 on 21 October 2026 (UTC+8). Their model IDs no longer work, so point any configuration to MiMo-V2.6-Pro. As of 25 October 2026, Xiaomi list MiMo-V2.6-Pro at the same $0.435 and $0.87 per million tokens, with 1M token context window too. Prices fit change, so read the current page before you plan budget.

The other option na your own hardware. Dem publish weights for MiMo-V2.5 and MiMo-V2.5-Pro on Hugging Face under MIT license. API retirement no remove the weights, so nothing stop you from serving dem yourself. But make you realistic about scale. Model wey get one million token context window no be work for ordinary VPS. You go need VPS wey get GPU attached or dedicated machine. Smaller local models na different kind work, closer to running Ollama on VPS to self-host LLM.

MiMo-Code fit also connect to other vendors. E fit migrate existing Claude Code login, connect OpenAI account, accept API key for another provider, or talk to any OpenAI-compatible endpoint. Check consumer subscription terms before you connect am. Provider terms decide which clients fit use subscription plan, and CLI from another vendor no always dey covered.

You suppose move comot from Claude Code?

If you already dey pay for Claude Code and e dey work, no. Version 0.1.15 for late September 2026, about three months after the first release and still the latest release as of 25 October 2026, na the honest signal. Version number wey be 0.1.x mean say releases fit introduce breaking changes, and configuration format still dey change. That one dey okay for tool wey you dey evaluate, but e fit cause pain when team dey depend on am.

MiMo-Code deserve serious consideration for four situations.

• Na you dey handle the bill. Xiaomi token prices far below frontier vendors, and the agent itself cost nothing.
• You need read the harness. Open source mean say you fit inspect prompts and tool definitions. That matter for security review or audit.
• You want buy the agent and model separately, so changing one no force you to change the other.
• Code no fit leave your network. Open weights make self-hosted backend possible instead of negotiation.

Cursor and Copilot na different comparison. Both dey inside editor, while MiMo-Code dey inside terminal. That difference affect more of your workday than any benchmark score, and comparison of Claude Code with Cursor, Codex and Copilot explain the trade-off in detail.

MiMo-Code run for VPS wey you don already rent

Terminal agent belong for server. E go hold session open for hours, e need stable network, and e suppose continue to run even if you close laptop lid. The normal way na the same method wey people use for other terminal agents: long-lived session for small server wey you fit reconnect to from anywhere, as how to run coding agent inside tmux for VPS explain. Once more than one agent dey work for that box, Claude Code fit send text directly from one session go another, so you no need move output by hand between tmux windows.

Give am dedicated unprivileged user, and point am to clone of the repository instead of your only copy. The access wey e get na the account wey e run as. Na this control still dey work if prompt go wrong.

The project dey distribute in two ways: shell installer wey dem serve over HTTPS, and npm package wey dem name @mimo-ai/cli for any machine wey already get Node.js. For server wey matter to you, prefer the npm package. You fit read wetin the package contain before e run, and you no dey pipe remote script straight into shell. For the bigger picture of other things wey deserve space for that box, the list of self-hosted AI agents na the map.

FAQ

MiMo-Code free?

The agent free and open source under MIT license, so you fit read and modify am. The model behind am na the part wey cost money. Xiaomi ship limited-time free channel wey dem call MiMo Auto at launch, and end MiMo Code free-use phase on 26 July 2026, so plan for paid access. Permanent price cut on 27 May 2026 set MiMo-V2.5-Pro at $0.435 per million input tokens when cache miss happen, and $0.87 per million output tokens. Xiaomi retire MiMo-V2.5-Pro on 21 October 2026, and dem list the successor, MiMo-V2.6-Pro, at the same rates as of 25 October 2026. You fit also point MiMo-Code to any OpenAI-compatible endpoint and pay another vendor instead.

MiMo-Code na the same thing as OpenCode?

E start as fork of OpenCode and e still share the core. The terminal interface, plugin system, provider abstraction, and MCP integration all come from that codebase. The repository LICENSE file still preserve the OpenCode copyright line. Xiaomi add persistent memory layer wey SQLite full-text search dey back, extra agent modes, and direct access to its own models. If you sabi OpenCode, you already sabi most of how MiMo-Code dey behave.

I fit use MiMo-Code with Claude or GPT models instead of Xiaomi own models?

Yes. First-run setup offer to migrate existing Claude Code login, connect OpenAI account, or accept API key for another provider. You fit configure any OpenAI-compatible endpoint by hand inside ~/.config/mimocode/mimocode.jsonc. Check consumer subscription terms before you connect am, because provider terms control which clients fit use subscription plan.

MiMo-Code safe to run on my main machine?

Treat am like any agent wey fit run shell commands. By default, e ask for approval before each action. --dangerously-skip-permissions turn those prompts off, and na how most accidents happen. Run am as dedicated unprivileged user, against clone of the repository, and preferably on machine wey you fit destroy and rebuild. The blast radius na the account wey e run as, so keep that account small.