Original source：Citi ResearchFirst published on Unhidea：Original source date：
Key takeaway
Citi believes M3.1 Flash illustrates a shift in model competition toward inference costs and integration into agent workflows. Continuity of overseas services has gained support, but the growth valuation still depends on customer retention amid competition and product execution.
More than 100 trillion tokens processed in 72 hours: what did Citi see?
MiniMax launched M3.1 Flash on 19 October and open-sourced the model weights on 23 October. The company disclosed that within 72 hours of launch, the model processed more than 100 trillion tokens for real users and developers worldwide. This is a model usage measure reflecting the scale of actual calls.
On that basis, Citi judged the model’s acceptance and usability to be high, potentially owing to better unit inference economics and improved practicality for enterprise-grade agent workflows. Its research focus consequently falls on whether tasks can be completed economically, rather than merely expanding parameter count.
What changed in efficiency and deployment compared with the previous generation?
M3.1 Flash has 428 billion total parameters and 23 billion activated parameters. Compared with the previous M3 model, the new version improves all-modal sparse attention, quantization-aware training and power-aware scheduling, with all three improvements directed toward inference efficiency.
The open-source release also came with deployment channels: MiniMax partnered with chipmakers including T-Head and Kunlunxin, and inference and cloud platforms including Together AI and Fireworks, to support open-source access on day one. It is also exploring cross-chip and cross-cloud applications with ecosystem participants such as vLLM and Hugging Face. The combination of efficiency optimization and deployment support underpins Citi’s view that competition is shifting toward end-to-end agent workflow integration.
How does progress on export registration affect overseas services?
A draft revision of the technology export catalog released for public comment on 16 October proposes adding AI technologies including large models and agents, with comments due by 15 November. From 16 October, exports of newly listed technologies require temporary registration until the revised catalog is finalized; registration itself does not constitute approval or authorization.
Citing media reports from 23 October and information from Yuyuan Tantian, the report said MiniMax had completed the registration and related procedures required to open-source its model and authorize overseas use, and could continue advancing the related business. Citi therefore believes continuity of services for global users has gained support.
What growth and discount assumptions underpin Citi’s target price?
Citi maintains its Buy/High Risk rating and a target price of HK$576, valuing the company at 7 times forecast 2028 revenue. The multiple applies a 30% discount to the historical average two-year forward price-to-sales multiple of 10 times; the target price also corresponds to 13.3 times forecast 2027 revenue.
The growth assumptions are high: Citi expects MiniMax’s revenue compound annual growth rate to be 160% over 2025–2030 and 120% over 2026–2029. The 30% valuation discount reflects the risks of intense competition, customers switching to alternative models, and a delayed or failed M3 Pro launch; the bank also lists the company’s ongoing funding needs as a constraint.
