Minimax stock jumps over 5% on new self-learning AI assistant

Xiyu Technology’s MINIMAX (00100.HK) introduced a cloud-based AI assistant, MaxHermes, that tackles the problem of skill adaptability in AI, sparking a 5.55% rally in its stock price. The new system introduces a “learning closed-loop” mechanism, a feature designed to allow the AI to autonomously learn and refine skills from user interactions, a direct challenge to more static, manually updated AI assistants.

The company’s announcement detailed that MaxHermes is built on the open-source Hermes Agent framework and will be deeply integrated with its own MiniMax M2.7 model. “After completing each task, the assistant automatically extracts and stores reusable skills, which can be loaded on demand in subsequent tasks and continuously refined based on new user feedback,” the company said in its release.

The market reacted positively to the strategic announcement, with MINIMAX shares closing up 5.55% to HK$895.00. The stock saw significant trading activity, with short-selling volume reaching $274.76 million, representing 14.7% of total turnover. The launch comes as the AI sector grapples with the high costs and architectural challenges of deploying increasingly powerful models.

This move positions MINIMAX to address what it calls a key “pain point” for traditional AI: a reliance on manually preset skills and a lack of long-term adaptability. By creating an assistant that improves with use, the company aims to deliver a more efficient and scalable AI solution that evolves alongside user needs, potentially lowering long-term maintenance and development costs.

Cloud vs. Edge: A Diverging AI Landscape

MINIMAX's cloud-based strategy for MaxHermes contrasts with another major trend in the industry: edge AI. As detailed in a recent report, many manufacturers are adopting on-site AI platforms running on ruggedized Industrial PCs (IPCs) to avoid the latency, security, and cost issues of the cloud. Companies like Emerson are championing this approach, where "milliseconds matter" and real-time feedback loops are essential for tasks like quality control on a factory floor.

The edge approach prioritizes low-latency, on-site processing for immediate decision-making, insulating operations from unreliable connectivity and keeping sensitive data in-house. However, cloud-based systems like MaxHermes offer distinct advantages in scalability, access to vast computational resources for complex learning tasks, and easier integration across a business. MINIMAX is wagering that for many enterprise applications, the power of a centralized, continuously learning brain outweighs the benefits of decentralized edge processing.

The 'Learning Loop' Advantage

The core of the MaxHermes announcement is its learning mechanism, which is based on the Hermes Agent, an open-source project from the U.S.-based firm Nous Research. Founded in 2023, Nous Research was already valued at approximately $1 billion as of April 2025, signaling strong investor confidence in its agent-based AI architecture.

Unlike models that require massive, periodic retraining on static datasets, an AI agent with a learning loop can dynamically acquire and improve specific "Skills." This could allow MaxHermes to become progressively more efficient at a user's specific tasks, from drafting reports to analyzing data, without waiting for a next-generation model release. This approach could significantly reduce the "knowledge decay" problem where models become outdated over time.

For investors, MINIMAX's strategy presents a compelling, if unproven, alternative to the dominant AI scaling paradigm. While competitors spend billions on GPU clusters for training ever-larger models, MINIMAX is focusing on software architecture to create a more efficient and adaptive product. The 5.55% stock increase reflects investor optimism that this "learning closed-loop" could become a key differentiator and a source of long-term value. If MaxHermes can deliver on its promise of continuous self-improvement, it could lower the total cost of ownership for enterprise clients and build a defensible moat based on accumulated, user-specific skills.

This article is for informational purposes only and does not constitute investment advice.