Xiaomi trained a frontier model with the garage door open.

30 steps of RL post-training, live: ≈1.5M rollouts, $3.47M of compute, every crash in public. This academy turns that raw dashboard into a path you can actually walk — 8 chapters, 5 simulators, a scored exam per chapter — from zero to reading telemetry like an engineer.

Begin chapter 01 · ~4 min Browse all chapters

Scroll · the run, decoded below

Total compute: $3,474,715
Attempts / step: ≈25,000
Pass rate: 0.50 → 0.64
Verdict: Headroom left; infra said stop

Mixture-of-Experts · 1T total / 42B active

mimo-v2.6-pro

stopped · step 30 · 2026-09-15 10:32 UTC

0.633 avg pass ▲0.068

cost$2,620,670

step3.43B

total75B

batch1,568×16

Smaller · faster · cheaper per step

mimo-v2.6-flash

stopped · step 30 · 2026-09-15 15:16 UTC

0.644 avg pass ▲0.130

cost$854,044

step3.7B

total81.4B

batch1,568×16

The two run cards, as Xiaomi showed them. Everything in this academy teaches you to read them.

◉ What you will learn

~40 min · 8 chapters · 5 labs · 8 exams

Each chapter has two depths: a plain-words story anyone can follow, and an engineer track with the real mechanism. Every chapter ends with a 10-question exam and a score. Walk them in order — they stack.

✦ Things that will make you curious

The run's most interesting facts — each one is a doorway into a chapter.

Why would a model edit the test instead of fixing its code?Reward hacking has three classic faces — and a chart-shaped fingerprint. → Chapter 04 Why is 100% success wasted compute?All-pass and all-fail tasks teach nothing. A live sampler routes around both. → Chapter 05 Why did a metric crash to zero overnight — twice?KL crashes mark restarts, not learning. Read the async heartbeat. → Chapter 06 Why does the 1T model take fewer tries but 2× longer?Fewer tokens, narrower pipe: the MoE hardware paradox. → Chapter 07 Why stop at step 30 with the curves still climbing?The last step carries a shutdown fingerprint, not convergence. → Chapter 08 What does every squiggle on the dashboard actually mean?16 instruments, each with rising / falling / watch-for verdicts. → Telemetry

▤ How the academy works

1 · Read a chapterPlain-words story first, engineer depth on demand — a toggle inside each chapter, not the homepage.

2 · Touch the labFive simulators live next to their concepts: run rollouts, crash clusters, solve the sampler's math.

3 · Take the exam10 questions, one at a time, instant feedback, score + miss review + retake. Clear 8/10 to feel done.

Start at the beginning Meet the simulators first