MiMo-V2.5 Goes Offline Tomorrow at 10:00: Your V2.6 Migration Checklist

MiMo-V2.5 Goes Offline Tomorrow at 10:00: Your V2.6 Migration Checklist (Oct 20, 2026)

One-line takeaway: Xiaomi's official notice: mimo-v2.5-pro / mimo-v2.5 retire on Oct 21, 2026 at 10:00 (UTC+8). If your code still references the old model names, spend 10 minutes switching today — requests will fail outright tomorrow.

What you need

  • Code, agent workflows, or tool configs calling the MiMo API with model set to mimo-v2.5-pro or mimo-v2.5
  • Ability to edit and fire one test request; run a regression pass before shipping to production

Step by step

Step 1: Read the official deprecation notice, memorize the mapping

Xiaomi published a formal deprecation notice in the MiMo open-platform docs. The timeline is unambiguous:

[Image to be added: 把下面这段 alt 文字当作图注 — Xiaomi MiMo official deprecation notice: V2.5 series retires Oct 21 10:00]

  • mimo-v2.5-pro → mimo-v2.6-pro (upgrade: the old one was text-only; the new one understands multimodal input)
  • mimo-v2.5 → mimo-v2.6-flash (watch the changed parameter defaults in Step 3)
  • Since 18:00 on Oct 14, requests to old names have been auto-routed to the new models. After 10:00 on Oct 21, the old names stop working entirely — requests error out.

Step 2: Grep for old model names and swap the IDs

Search your projects for mimo-v2.5:

# find every reference to the old models
grep -rn "mimo-v2.5" --include="*.py" --include="*.ts" --include="*.js" --include="*.json" --include="*.yaml" .

Replacement map:

Old (dead tomorrow) New (switch to)
mimo-v2.5-pro mimo-v2.6-pro
mimo-v2.5 mimo-v2.6-flash

Calling via OpenRouter? The new model IDs there are Xiaomi: MiMo-V2.6-Pro (and the Flash counterpart), priced the same as the official API — Pro at $0.435/$0.87 per million input/output tokens, Flash at $0.14/$0.28. Xiaomi says V2.6 keeps previous-gen pricing.

[Image to be added: 把下面这段 alt 文字当作图注 — Xiaomi models on OpenRouter: V2.6 pricing and model list]

Step 3: Check three parameter changes — don't just swap names

The official notice documents V2.6 behavior changes. A bare model-name swap can silently change behavior:

  1. No more custom temperature / top_p in thinking mode: v2.6-flash and v2.6-pro force temperature: 1.0 and top_p: 0.95 — anything you pass is ignored. If you tuned sampling params before, delete or comment out that logic.
  2. max_completion_tokens default changed: mimo-v2.5 defaulted to 32768; mimo-v2.6-flash defaults to 131072. If you relied on the default for truncation/pagination, outputs can now be 4× longer — either pass the param explicitly or check downstream handling.
  3. Multimodal is new: mimo-v2.6-pro accepts image input. If you previously routed around v2.5-pro's text-only limit (e.g., calling a separate vision model), you can simplify the chain.

Step 4: Regression-test these three things before production

  • Long-output tasks: with the bigger default max tokens, check long generations don't blow up downstream consumers
  • Multi-step agent tasks: third-party benchmarks scored v2.5-pro 87.5% average on agent tasks (97.9% on coding); Xiaomi claims stronger reasoning for v2.6 — verify with your own task set instead of assuming "newer is better"

[Image to be added: 把下面这段 alt 文字当作图注 — Third-party benchmark: mimo-v2.5-pro agent-task scores]

  • Billing: prices are unchanged, but longer outputs = more tokens = bigger bills. Glance at your usage dashboard.

Notes, costs, pitfalls

  • The deadline is tomorrow 10:00 (UTC+8): auto-routing since Oct 14 was only a transition. After that, old names hard-fail. Cron jobs and scheduled tasks with hardcoded model names are the easiest to miss — grep everything (Step 1's command).
  • Same price ≠ same bill: unit prices didn't change, but the 4× larger default max tokens means more output tokens per call. Cost-sensitive workloads should pass max_completion_tokens explicitly.
  • Silent thinking-param failure is the sneakiest: passing temperature/top_p doesn't error — it's just ignored, so behavior differs from what you expect. Check this first when debugging.
  • OpenRouter users: model names carry the Xiaomi: prefix — don't confuse with the direct-API names mimo-v2.6-pro. OpenRouter also makes it easy to A/B against other models, but watch cache hit rates — they materially affect your real bill.

Sources

  • Xiaomi MiMo official deprecation notice (migration map + parameter changes): https://mimo.mi.com/docs/en-US/updates/deprecate
  • Xiaomi models on OpenRouter (V2.6 pricing and model list): https://openrouter.ai/provider/xiaomi
  • Third-party benchmark (mimo-v2.5-pro agent-task scores): https://kilo.ai/models/xiaomi-mimo-v2-5-pro
  • Sunset news report (Oct 12, 2026): https://www.binance.com/fr-AF/square/post/376356828991380

评论

此博客中的热门博文

Oracle Cloud Free Tier: Spinning Up the 2-Core 12GB Ampere A1 Instance

Wise Multi-Currency Account Walkthrough: 40+ Currencies + Local Bank Details