Xiaomi MiMo v2.6 Flash: An Assessment on Two Real Engineering Tasks

Cover image: Xiaomi MiMo AI Model — Introduction and hands-on assessment

Xiaomi’s MiMo is a large model I drive through the OpenCode agent or directly via the Xiaomi platform API (where I bought tokens for v2.6). I’d used the free OpenCode/MiMo v2.5 for months to draft blogs reliably; with v2.6 out, I bought paid API tokens and stress-tested the model on two real, hours-long engineering tasks.

Task 1 — End-to-end feature

Major features shipped and passed manual functional testing with no functional issues. After 3 automated review rounds (my hard cap), 4 minor issues remained.

End-to-end feature — final agent report: what was built (DB migration, IPC group sync, capture, UI group filter, tests) + green verification table + R1–R3 review log

End-to-end feature — compliance checklist: stopped at 3-round bound, PR left unmerged, pre-existing master failures disclosed

End-to-end feature — manual test guide: running against the real DB with the new migration applied

Cost: ¥1.96, 239 requests, 27,925,352 tokens (mimo-v2.6-flash).

End-to-end feature — usage detail: total ¥1.96

End-to-end feature — request count: 239, all mimo-v2.6-flash

End-to-end feature — token usage: 27,925,352 tokens

Task 2 — App release

Major requirements met. The one miss: the release-note text was wrong — the note’s git tag was intact, but a downstream publish step never read it back, and the 3-round review didn’t catch it.

App release — CI result: pipeline #7 green, release-* correctly skipped for -rc, 6 assets + tag v0.1.0-rc2 delivered

App release — release-note post-mortem: tag intact, downstream never read it back, 3-round review missed it

Cost: ¥0.68, 83 requests (mostly mimo-v2.6-flash, some mimo-v2.6-pro), 9,331,424 tokens.

App release — usage detail: total ¥0.68

App release — request count: 83, mostly mimo-v2.6-flash + some v2.6-pro

App release — token usage: 9,331,424 tokens

Assessment

MiMo handles hours-long engineering tasks — both an end-to-end feature and an app release were largely completed on v2.6 Flash.
The free tier is fine to start, but long tasks are risky — the free quota has a daily cap; a hours-long task can halt mid-run with “quota exhausted.” That is exactly why heavy tasks need paid API.
Tune the prompt, skill, and workflow — the residual minors (and the release-note slip) show the setup still needs tuning to avoid these and other limits, not just a human cleanup pass.

Both tasks finished at acceptable cost; the residual minor issues still need follow-up assessment.