Xiaomi open-sourced MiMo Code and made MiMo V2.5 free for a limited time: perk or developer-entry war?

I received a Xiaomi MiMo email today. The phrase “open-sourced” did not stop me. Open-source AI tools are everywhere now: forks, wrappers, new workflows, and small UX differences. What stopped me was another line in the email: MiMo Code includes the top-tier multimodal model MiMo V2.5 for free, for a limited time.

The important part is not “free.” It is “for a limited time.”

In developer tools, those four words are rarely just a simple perk. They usually point to three things: workflow entry, quota, and future billing.

MiMo Code is a terminal-native AI coding assistant from Xiaomi’s MiMo team. Xiaomi says it is built on top of OpenCode, open-sourced under the MIT license, and designed for long-horizon programming tasks where an agent must maintain state and decision quality across dozens or hundreds of steps. On first launch, users can choose MiMo Auto, Xiaomi MiMo platform login, imported Claude Code configuration, or a custom model provider. MiMo Auto is the notable part: it is free for a limited time, based on MiMo V2.5, and supports a one-million-token context window.

If Xiaomi had only open-sourced a terminal agent, that would be interesting but not a new story. The more important move is that Xiaomi is putting a model entry point directly into a coding workflow and making it free first.

MiMo Code is not an isolated release. Xiaomi has been building this line for a while. MiMo-7B focused on math and code reasoning. MiMo-V2-Flash moved toward a larger MoE model built for reasoning, coding, and agentic workflows. MiMo-V2-Pro drew outside attention when the anonymous Hunter Alpha model on OpenRouter was later reported to be an early test version from Xiaomi. MiMo-V2.5 and MiMo-V2.5-Pro then pushed the line further into long-context, multimodal, and agentic coding scenarios.

So I read MiMo Code as part of a sequence: Xiaomi builds models, tests agent scenarios, offers Token Plan billing, and now open-sources a tool that can become a developer workflow entry point.

There are several disputes here.

First, is this an open-source tool or an entry-point strategy? MiMo Code is a fork built on OpenCode, and that alone is not a problem. Forking is normal in open source. The real question is whether the fork adds value. Xiaomi’s additions target real pain points in coding agents: persistent project memory, context management, subagent orchestration, goal loops, Compose workflows, and Dream/Distill mechanisms for maintaining and distilling experience.

This is engineering work, not just a slogan. But marketing phrases such as “near-infinite context” and “self-evolution” can mislead normal readers. The model does not literally remember everything forever, and its weights are not rewriting themselves as you use it. A more grounded translation is: the tool stores project-level memory, rebuilds state when context gets long, and periodically cleans or distills past sessions into reusable project knowledge and workflows.

Second, the limited-time free access is the real hook. Free access is good, but as someone who manages software costs in a small company, I do not only ask whether I can use it today. I ask what the bill will look like later.

Xiaomi’s MiMo Open Platform already shows the commercial path. Token Plan is positioned for AI programming scenarios, covers MiMo V2.5 models and other model services, and integrates with tools such as OpenCode, OpenClaw, and Claude Code. The platform has also adjusted MiMo V2.5 pricing and Token Plan quota. This is not a conspiracy. Large model providers need revenue, and coding agents burn tokens quickly because they read files, edit code, run commands, inspect errors, and iterate.

From a business perspective, the sequence makes sense: lower the barrier with limited-time free MiMo V2.5 access, capture the terminal workflow with an open-source tool, and then keep serious users within MiMo Platform, Token Plan, or at least the MiMo developer ecosystem.

That can be both a real perk and a developer-entry war.

Third, will domestic developers actually accept this path? In China, many developers are not avoiding Claude Code or Codex because they dislike them. They often face network, payment, account-risk, reimbursement, and compliance problems. For small teams, “globally strongest” is not always the most useful attribute. Stable access, RMB billing, invoices, and support for domestic model providers can matter more.

DeepSeek is the practical comparison. DeepSeek documents integration with AI coding tools such as Claude Code, OpenCode, and OpenClaw. Its V4 Flash and V4 Pro models support one-million-token context, tool calls, JSON output, and token-based API pricing. This means MiMo Code’s real competition in China may not be Claude Code itself. It may be the question of who becomes the shell for domestic coding agents.

That shell can run MiMo, DeepSeek, Kimi, GLM, or other providers. Users may arrive for free MiMo V2.5, but if the tool is useful, they may stay in the workflow even after switching models.

Early community feedback should not be overstated. Some users question “near-infinite context.” Others discuss MiMo Auto’s free boundary, reverse proxy possibilities, Token Plan friction, 500 errors, 429 errors, and latency. These are not enough to declare success or failure. But they reveal a very real domestic developer mindset: when a free model appears, people do not first ask whether they are loyal to the brand. They ask whether it can be used for free, whether it is stable, and what happens after the free period.

That is the real test for Xiaomi.

AI coding tools are not phone launches. Parameters, benchmarks, and limited-time free access can bring attention, but developers stay for stability, controllability, transparent billing, safe permissions, and predictable behavior in real projects.

MiMo Code can read and write code, run commands, and manage Git. The more powerful a terminal agent becomes, the more carefully it should be tested. I would try it first on toy projects, small tools, and non-sensitive repositories. I would not put a company’s main codebase into a new agent on day one.

My current judgment is simple: Xiaomi did not merely release a domestic Claude Code alternative. It is using an open-source terminal agent to connect MiMo models, limited-time free access, Token Plan billing, domestic model integrations, and developer workflow entry into one line.

This line has real engineering value and real marketing taste. It is a perk, and it is also customer acquisition. It is open-source, and it is also a commercial entry point.

The question is not whether MiMo V2.5 is free today. The real question is whether developers will still be willing to hand Xiaomi their projects, time, and bills after the free period ends.

老花 / Easton Hua