M2.5 detonates Agent demand! MINIMAX-WP(00100) has the highest cumulative increase of over 480% on the market.

date
12:03 24/02/2026
avatar
GMT Eight
In the Lunar Year of the Horse in 2026, MiniMax-WP (00100) has undoubtedly become one of the most eye-catching focus in the Hong Kong stock market.
The Lunar New Year of the Horse in 2026 undoubtedly became one of the most attention-grabbing focal points in the Hong Kong stock market for MiniMax-WP (00100). On February 20th, the first trading day of the Lunar New Year of the Horse in Hong Kong, MiniMax's stock price closed up 14.52%, with a market value surging above 304.2 billion Hong Kong dollars, reaching a historical high since its listing. Since its debut on the Hong Kong Stock Exchange on January 9th at an IPO price of 165 Hong Kong dollars per share, it has achieved a cumulative increase of over 480%, successfully establishing itself as a core target in the Hong Kong stock AI industry. This is not simply driven by market sentiment, but rather by the industry-level eruption triggered by the release of its new-generation model M2.5 after the Spring Festival. According to OpenRouter platform data, within 12 hours of the release of M2.5, it topped the OpenRouter popularity chart, and within a week, it topped the call volume chart, with the weekly call volume skyrocketing to 3.07T tokens, surpassing the total of Kimi K2.5, GLM-5, and DeepSeek V3.2. Behind the call volume is the collective awakening of the Agent ecosystem. The eruption of M2.5 is not only due to its price advantage (input price of 0.103 US dollars / million tokens, output price of 1.34 US dollars / million tokens, only 1/20 of Claude Opus 4.6), but also because it precisely broke through the "cost-ability" critical point of the Agent workflow. OpenRouter officials disclosed that M2.5 leads in the call volumes of long texts in the 100K-1M token range, which is the most representative consumption range in the Agent workflow. At the technical level, MiniMax has restructured AI execution logic with "Forge." Traditional models face two major pain points in Agent scenarios: context management relies on external rules leading to logical breaks, and training efficiency is hindered by repetitive prefixes. MiniMax has achieved a fundamental breakthrough with its self-developed Forge system: Architecture decoupling: Gateway Server isolates Agent behaviors and model complexity, Data Pool asynchronously collects training trajectories; Training acceleration: Prefix Tree Merging reconstructs training samples into a tree structure, achieving approximately 40 times faster training speeds; Engineering optimization: Windowed FIFO strategy balances throughput efficiency and stability, and the compound reward mechanism (process reward + time reward) guides the model to choose the optimal execution path. This design directly empowers top open-source tools. Core frameworks such as Kilo Code and OpenClaw have fully integrated M2.5, with an 80.2% pass rate in SWE-Bench Verified evaluations and ranking first in the Multi-SWE-Bench task in multiple languages. The rise of AgentOS is rewriting the rules of competition for large models. When token consumption shifts from "text interaction cost" to "action execution cost," model manufacturers must shift from "single-point capability competition" to "system-level adaptation." The eruption of MiniMax M2.5 confirms this trend: it not only meets developers' demands for "functionality and affordability," but also deeply integrates models into the Agent workflow through the Forge system, enabling the execution of complex tasks to move from demos to scale deployment. Currently, the strong rise in the stock price of MiniMax is actually a recognition of the value of "AI execution layer infrastructure" by the market. Behind the 3T call volume of M2.5 on OpenRouter is the vote cast by Silicon Valley developers based on real needs - as the Agent ecosystem transitions from a concept to reality, MiniMax is defining the underlying logic of the next generation of AI applications as a technological innovator. As pointed out by the a16z research institute, the true competitiveness of a model lies in the "long-term fit of the workload," and MiniMax's Forge architecture has already laid the key foundation for this transformation.