Goldman Sachs maintained its Buy rating on Xiaomi (1810.HK) with a 12-month target price of HKD39, citing the MiMo-V2.6 reinforcement-learning livestream. The broker said MiMo-V2.6-Flash gained more than 19 DeepSWE points in 3.5 days for USD850,000, versus USD2.62 million for Pro's gain of more than 14 points. Both models stopped after 30 training steps, in what Goldman called the first real-time public disclosure of post-training telemetry by a major AI lab.
Goldman Sachs kept its Buy rating on Xiaomi (1810.HK) and a 12-month target price of HKD39, after the company's MiMo-V2.6 reinforcement-learning run showed the smaller Flash model delivering more benchmark gain per dollar than the trillion-parameter Pro.
The broker said MiMo-V2.6-Flash added more than 19 points on DeepSWE over 3.5 days for a total spend of USD850,000, working out to USD10 per million tokens, or USD2.85 per second. MiMo-V2.6-Pro spent USD2.62 million, or USD35 per million tokens and USD5.71 per second, for a gain of more than 14 points — a cost rate two to three times higher than Flash.
Both models stopped after 30 training steps. Goldman said the exercise marked the first time a major AI laboratory had publicly disclosed post-training telemetry data in real time, and that the RL work had materially improved performance on long-horizon agent tasks.
The disclosure gives the market its first dollar-denominated yardstick for post-training spend at a Chinese frontier lab. Xiaomi's own release materials put Flash at roughly 309 billion total parameters with 15 billion active per token, against Pro's 1.02 trillion total and 42 billion active — a sparse mixture-of-experts design in both cases, with a context window of up to 1 million tokens and text, image, video and audio input. Xiaomi has published the weights under an MIT license alongside more than 7,000 RL task environments and an end-to-end training framework, a combination that lets outside teams test whether the cost-per-point advantage survives replication.
The efficiency gap is the crux of the Goldman case. If a 15-billion-active-parameter model can recover most of a flagship's agent capability at roughly one-third the training cost, the economics of serving long-horizon coding and tool-use workloads shift toward the cheaper tier — the same pattern Google has pursued with its Gemini Flash line and DeepSeek with V4.1 Flash. Xiaomi's API pricing reflects that split: Flash lists at $0.14 per million input tokens and $0.28 per million output tokens, against $0.435 and $0.87 for Pro.
Positioning cuts the other way. Short selling in Xiaomi stood at USD811.36 million as of 2026-09-21, a short ratio of 21.652%, according to data cited alongside the note. That is a heavy bearish book for a stock carrying a Buy rating and a target implying upside from current levels, and it raises the odds of sharp moves on any AI-related headline in either direction.
The read-through extends past Xiaomi. Goldman's USD10 versus USD35 per million token comparison lands as a sector-level cost benchmark, arriving in the same week that Chinese and US labs traded claims over frontier model leadership and analysts flagged a capability gap in humanoid robotics between the two markets. For Xiaomi holders, the next test is whether the MiMo team open-sources the technical details it has promised, which would let independent researchers verify the cost figures rather than take the dashboard at face value.
This article is for informational purposes only and does not constitute investment advice.
