More than 100 trillion tokens processed in 72 hours: what did Citi see?

  • MiniMax launched M3.1 Flash on 19 October and released the model weights on 23 October. The company disclosed that, within 72 hours of launch, the model had processed more than 100 trillion tokens for real users and developers worldwide.
  • Citi sees this scale of usage as a sign that the model is easy to adopt and practical. It believes the reasons may be lower inference cost per unit and improved utility in enterprise-grade agent workflows. Its focus has therefore shifted from how large the model is to whether it can be integrated into real tasks at a lower cost.

How have efficiency and access changed from the previous generation?

  • M3.1 Flash has 428 billion total parameters and 23 billion activated parameters. Compared with the earlier M3 model, sparse attention, quantization-aware training and power-aware scheduling have all improved. The report directly links these changes to higher inference efficiency.
  • The open-source release also comes with chip and cloud support: MiniMax has partnered with chipmakers including T-Head and Kunlunxin, as well as inference and cloud service platforms including Together AI and Fireworks, to support open access from day one.
  • The company is also working with inference engines and open-source communities such as vLLM and Hugging Face to explore applications across chips and cloud platforms. The business connection here is deployment and developer access: Citi values the combined contribution of efficiency improvements and ecosystem support to model adoption.

What uncertainty did completion of the overseas procedures ease?

  • Proposed revisions to the technology export catalogue announced on 16 October cover AI technologies including large models and agents, with consultation open until 15 November. Temporary registration requirements apply to newly added technologies from 16 October; the final catalogue has yet to be determined. Registration itself is a separate procedure from approval and authorization.
  • Citing media reports from 23 October, the report says MiniMax has completed the registration and related procedures required to open-source the model and authorize its use overseas, and can continue the related work. On this basis, Citi believes continuity of the company’s services to users worldwide has been secured.

Why maintain a Buy rating while still applying a thirty percent valuation discount?

  • Citi maintains its Buy/High Risk rating and a target price of HK$576, valuing the company at 7 times its forecast 2028 revenue. That is a 30% discount to the historical two-year average forward price-to-sales multiple of 10 times used in the report.
  • The discount reflects three specific constraints: intense competition, potentially significant volatility if customers switch to other models, and a possible delay or failure in the launch of M3 Pro. Flash’s adoption speed and efficiency improvements support the assessment of the product, while customer retention and the performance of subsequent new products continue to constrain valuation.
  • Citi also lists ongoing financing needs, a short trading history since listing and the rapidly changing AI environment as risks, and therefore retains both its positive product assessment and its High Risk rating.