If you've been following the AI tooling series here — the Claude Code one-month retrospective, the OpenCode verdict, the harness explainer — you know the pattern by now. I don't review a model off a benchmark chart. I point it at real work on my actual stack (AlmaLinux VPS, Podman, Caddy, the whole terminal-only setup) and see what survives contact with production.

This time it's Xiaomi's MiMo-V2.5. Open-source, MIT licensed, and — on paper — punching well above what a 15B-active-parameter model has any right to punch. I ran it through OpenCode via BYOK for a month across six use cases: DevOps automation, pentest workflows, code review, vibe coding, full implementation tasks, and deep research. Here's what actually happened.

What It Actually Is

MiMo-V2.5 is a 310B-parameter sparse Mixture-of-Experts model with only 15B active per inference pass. That gap between total and active parameters is the whole story — you get frontier-adjacent capability without frontier-tier compute bills.

The architecture leans on hybrid sliding-window attention: local SWA layers interleaved with global attention layers at a 6:1 ratio, which cuts KV-cache storage to roughly a seventh of what full attention would need. That's what makes the 1M-token context window survivable to actually serve, instead of just a number on a spec sheet. It's trained on 48T tokens across five stages, including agentic post-training where the context window gets progressively stretched from 32K to 256K to the full 1M.

It's also natively multimodal — text, image, video, and audio in one model, no routing between specialists. And it ships under MIT, so no authorization dance for commercial use, fine-tuning, or self-hosted deployment.

How I Actually Ran It

A quick note on setup, because it matters for how you read the rest of this post.

MiMo-V2.5 is available for free directly in OpenCode, under Zen. That's a genuinely nice on-ramp — but read the fine print before you get comfortable. Most models on Zen run zero-retention, no training, full stop. MiMo-V2.5 Free is explicitly called out as an exception: while it's free, OpenCode's own docs state that collected data may be used to improve the model. Fair enough — Xiaomi's earning feedback in exchange for free compute — but it means the free tier is not the place to run anything touching client code, credentials, or infrastructure you can't afford to have inspected. I kept free-tier sessions to disposable test repos and public code only.

For the real month-long test, I paid for it. I subscribed to Xiaomi's Lite tier on their own Token Plan for a month — the cheapest paid rung, well under $10 — specifically so I could push it under normal usage terms instead of "free preview" terms. I was also already running an OpenCode Go subscription before this review started, which happens to be the cheapest way I've found to get consistent access to both MiMo-V2.5 and DeepSeek V4 side by side without juggling separate API keys. Go is flat-rate, dollar-metered usage against a curated model list, and MiMo-V2.5 is one of the least expensive seats at that table — which made it an easy model to lean on heavily for a month without worrying about the bill.

The Math

Free via OpenCode Zen: genuinely free, but not private — that free-tier data exception is real, so treat it as fair-use territory for public code and experiments, not client work.

Paid, the way I actually ran it: Xiaomi's own Lite tier on the Token Plan for solo testing, and OpenCode Go for the day-to-day grind — the cheapest route I've found to keep MiMo-V2.5 and DeepSeek V4 both on tap without separate billing relationships.

API pricing: MiMo-V2.5 runs at 1x credits, MiMo-V2.5-Pro at 2x — and Xiaomi doesn't charge a separate multiplier for the 1M context window, which used to be the sneaky cost trap with long-context models.

Compared to alternatives: Xiaomi's own benchmarks claim 40–60% fewer tokens than Claude Opus 4.6 and GPT-5.4 for comparable agentic task completion — which, in practice, translates into a real cost advantage per trajectory, even if it's not a clean 1:1 price comparison. The 15B active parameter count is also why self-hosting economics look so different from models with heavier active counts.

The Orbit Program: Xiaomi ran a limited-time launch promo distributing 100 trillion tokens over 30 days, with approved applicants getting up to 1.6 billion credits (worth roughly 659 yuan). That window has since closed, but it's worth knowing it happened — it's part of why early adopters got such a favorable cost picture.

The Verdict

Reach for MiMo-V2.5 if:

  • You need long-context reasoning across large codebases or document sets
  • You want an agentic model that uses tools well without burning excess tokens
  • You're cost-sensitive and want frontier-adjacent capability without frontier pricing
  • You need multimodal understanding — text, image, audio — in one model
  • Vendor lock-in is a dealbreaker and MIT licensing actually matters to your stack

Skip it if:

  • You need near-zero hallucination on critical-path work (Opus-tier models still hold the edge there)
  • You're doing serious visual reasoning — dedicated vision-first models are still ahead
  • You want a model that magically knows every library released last week (no model does)
  • You just want a simple chat model without agent/tool-use complexity

Bottom line: MiMo-V2.5 isn't the flashiest model I've run through OpenCode this year, but it's become the one I reach for most by default. The combination of real 1M context, native multimodal, solid agentic behavior, and MIT licensing makes it hard to argue with on value alone. For DevOps, code review, and research work specifically, it's earned a permanent spot in the rotation.

It won't replace a senior engineer. But paired with the right harness, it'll make one meaningfully faster — and that's the metric that actually matters.