Since 15 September Xiaomi has run the reinforcement-learning phase of its next model on a public dashboard. Mid-run, MiMo-V2.6-Pro scores 65.97% on the DeepSWE v1.1 agentic coding benchmark against 19% for the shipped MiMo-V2.5 — fourth among frontier models, behind GPT-6 Astra, Claude Fable 5 and Kimi K3, on Forkast's reading of the same page. Xiaomi's accounts carry no AI revenue line, no AI margin and no AI segment result, so an investor has nothing to attach a price to.
Seventeen months from 7 billion parameters to a trillion
Fuli Luo, who leads the MiMo team, opened the run to the public after what his post calls "nearly half a year of silence" spent on one question: how far reinforcement learning scales. Each step processes about 2 billion tokens across 1,568 prompts at sixteen asynchronous rollouts apiece, with code, general, visual, cybersecurity and chat-agent tasks trained together in one harness rather than a specialist run per domain. Two runs are live, Pro and Flash.
Open-weight releases by Xiaomi's MiMo team — company releases, not R40 estimates
| Model | Released | Parameters |
|---|---|---|
| MiMo-7B | April 2025 | 7B |
| MiMo-V2-Flash | December 2025 | 310B (15B active) |
| MiMo-V2-Pro | March 2026 | 1T (42B active) |
| MiMo-V2.5 | April 2026 | 311B |
| MiMo-V2.5-Pro | April 2026 | 1T |
| MiMo-V2.6-Pro | In RL since 15 Sep 2026 | Not disclosed |
Release dates are from the published MiMo release history; parameter counts are those listed on the open weights, with the sparsely activated count in brackets where the team publishes one. MiMo-V2.6 has no release, no published parameter count and no benchmark of its own — its reinforcement-learning run has been visible on Xiaomi's live dashboard since 15 September 2026, and the 65.97% DeepSWE v1.1 score quoted in the article is a mid-run reading of that dashboard, not a shipped model's result.
Alain Lam, Xiaomi's finance chief, had already told analysts this was coming. On the 18 August results call: "you will pay close attention to our new model launch. It is being trained now. Very soon, it will face the world." Four weeks later anyone can watch the reward curves.
The whole business sits in a RMB1.0bn residual
Xiaomi's June quarter reported RMB24.9bn in the Smart EV, AI and other new initiatives segment. Smart EV was RMB23.9bn of that. Every other business in the segment — models, robotics, the rest — shares the remaining RMB1.0bn, and Lam said on the call that is where it stays: "now MiMo just started to be monetized and the revenue is not big. In the future, at appropriate times, we may disclose it separately. But at this stage, we will still put it within this RMB 1 billion."
RMB1.0bn ÷ RMB108.9bn of group revenue = 0.92%. Read it as a ceiling, not a measurement: robotics and everything else live in the same line, so MiMo's share is smaller by an amount nobody outside Xiaomi knows.
Spending is disclosed once, and as a share. First-half R&D was RMB18.2bn, up 25.6%, "with AI-related investments accounting for nearly 30%," Lam said. RMB18.2bn × 30% = RMB5.46bn across six months, so RMB2.73bn a quarter, or 2.7 times the entire line the revenue is reported inside. "Nearly 30%" is the softest number in that chain, so here it is three ways:
| AI share of first-half R&D | AI spend per quarter | Against the RMB1.0bn line |
|---|---|---|
| 25% | RMB2.28bn | 2.3x |
| 30%, the company's wording | RMB2.73bn | 2.7x |
| 35% | RMB3.19bn | 3.2x |
The live cost counter is a rate, not a bill
One running number about the programme is public, and it measures nothing. The dashboard's client code computes total cost as a fixed rate times wall-clock elapsed — $5.71 a second for Pro, $2.855 for Flash — which is why two write-ups of the same run on the same day put the Pro burn at $432,000 a day and at $493,000. Xiaomi publishes the rate and no hardware, so nothing converts it into a GPU bill.
Nobody publishes a target on the line a US reader sees
XIACY is an unsponsored ADR. Xiaomi neither created it nor files with the SEC, and US brokers carry no rating on it; the published targets are set on the Hong Kong ordinary share, 1810, in Hong Kong dollars, five of which make one ADR. Whatever the sell side thinks of MiMo, no consensus figure exists against the ticker a US reader is quoted.
The strongest case against all this is that the market prices the programme implicitly, inside the multiple on the whole company. Take that at face value and the multiple has gone the other way. At the $17.64 close on 20 August, two days after the results, 5.30bn diluted ADS-equivalents were worth $93.5bn; add $0.30bn of net debt, divide by trailing-four-quarter revenue of $62.6bn, and the answer is 1.50 times. At $16.62 on 17 September, $88.1bn and 1.41 times. Same revenue base in both, so the whole move is the share price, down 1.6% from the 15 September close of $16.89.
Lam said Xiaomi may separate AI revenue "at appropriate times." The September-quarter results are where that happens or does not: a disclosed AI line, at any size, gives the market something to price and ends the argument made here.
Benchmark scores, step size, prompt count, rollouts and the run start date are Forkast's reading of Xiaomi's live dashboard; the cost rates and the counter's construction are ours, read off the page itself on 18 September, and it has moved since. The $432,000 and $493,000 daily figures are two third-party write-ups of the same run. Release dates and parameter counts come from the published MiMo release history and the open weights. Segment revenue, group revenue and R&D are Xiaomi's 18 August results announcement; the quotes are from that day's investor call, spoken through an interpreter. Every share, multiple and quarterly split here is ours. Dollar revenue converts the filed renminbi at each quarter's average rate — Xiaomi publishes no dollar translation. Prices are the 20 August, 15 September and 17 September closes.
