April 28, 2026
Bot codes, comments explode
Xiaomi releases MiMo-v2.5 Family weights with strong coding and agent benchmarks
Open‑source Xiaomi AI aces coding tests — fans cheer, skeptics cry “slop”
TLDR: Xiaomi open‑sourced MiMo‑V2.5‑Pro, an AI that rapidly built real software and scored near top coding models. Commenters are split between excitement over the free MIT release and harsh takedowns of hypey coverage, with lingering questions about pricing, practicality, and whether MiMo is a true challenger.
Xiaomi just dropped its MiMo‑V2.5‑Pro AI as open source (MIT license), and the comment sections immediately lit up. The company claims this bot blitzed a full college‑level compiler in 4.3 hours, built a basic video editor overnight, and even tuned a tricky chip design in about an hour. Fans are buzzing that it runs with top models on coding tests and can plan and fix itself over thousands of steps — the kind of “agent” behavior people drool over. The receipts? Xiaomi’s official page and benchmark charts.
But the drama is the headline. A chunk of the crowd is roasting the breathless write‑ups — one user groaned “Ugh, the slop,” while another linked a more sober VentureBeat piece. Curious onlookers keep asking if MiMo is a serious contender or just shiny demo magic, and the practical folks want to know what it costs to use: Xiaomi prices “credits,” not tokens, and that’s confusing. Meanwhile, hype lovers are thrilled it’s MIT Licensed and reportedly on Hugging Face, meaning anyone can tinker.
The memes wrote themselves: “TA’s replaced by Tuesday,” “Built a video editor while we slept,” and “students in shambles.” Whether MiMo is the next coding coworker or just another benchmark banger, the community’s split — half ready to clone the repo, half reaching for the eye bleach.
Key Points
- •Xiaomi open-sourced MiMo‑V2.5‑Pro, an upgrade over MiMo‑V2‑Flash with improved long‑horizon and agentic capabilities.
- •MiMo‑V2.5‑Pro completed a Rust SysY compiler in 4.3 hours, passing 233/233 hidden tests from a Peking University project and self‑correcting mid‑process.
- •The model built a functional desktop video editor in 11.5 hours using 1,868 tool calls, producing 8,192 lines of code and standard editing features.
- •In an analog design task (FVF LDO on TSMC 180nm) integrated with ngspice, the model met targets in ~1 hour, improving line and load regulation 22× and 17×.
- •Benchmarks show MiMo‑V2.5‑Pro near top models on SWE‑Bench Pro (57.2), leading Opus 4.6 on Terminal‑Bench 2.0, and outperforming GPT‑5.4 and Gemini on Claw‑Eval.