MiniMax market capitalization approaches 200 billion HKD! AI "game-changing" application arrives

On the eve of the Year of the Horse Spring Festival, the global AI community welcomes the “Battle of the Gods.”

On the afternoon of February 12, MiniMax launched its latest flagship programming model MiniMax M2.5, becoming the world’s first production-level model designed natively for agent (intelligent agent) scenarios, with programming and agent performance comparable to top international models.

On February 6, Anthropic and OpenAI respectively released new versions of their foundational large models Claude Opus 4.6 and GPT-5.3-Codex.

In the early hours of February 12, Zhipu open-sourced the new generation base model GLM-5.

During trading on February 12, both MiniMax and Zhipu saw significant gains, with MiniMax’s market value approaching 200 billion Hong Kong dollars at one point. By the close of trading on February 12, MiniMax’s stock price was 588 HKD per share, up 14.62%, with a total market capitalization of 184.4 billion HKD.

MiniMax M2.5 Competes with Claude Opus 4.6

MiniMax M2.5 is comparable to Claude Opus 4.6, supporting full-stack programming development across PC, app, and cross-platform applications. It leads the industry in high-level Excel processing, in-depth research, PPT, and other core productivity scenarios.

Claude Opus 4.6 was released by Anthropic on February 6. It is reported to sustain longer autonomous workflows and surpass competitors like GPT-5.2 in key enterprise benchmarks.

Technical highlights of MiniMax M2.5 include flagship-level programming and agent performance, native agent architecture, the next-generation main model for digital office, advantages in private deployment, and extreme reasoning efficiency.

For example, in private deployment, MiniMax M2.5 has only 10 billion parameters, making it the smallest flagship model among the first tier of AI large models. It has advantages in private deployment, memory usage, and inference efficiency.

Additionally, MiniMax M2.5 has deeply optimized the “thinking” process, supporting ultra-high throughput of 100 TPS (transactions per second), with inference speed three times that of Claude Opus 4.6.

AI Community Competing Fiercely, MiniMax M2.5 Excels in Industry Cost-Performance

Recently, the AI community has seen a new product showdown. Domestically, Zhipu open-sourced the new generation base model GLM-5, marking the beginning of an era where large model programming enters a complete system engineering phase for agent engineering.

Abroad, Anthropic launched Claude Opus 4.6, and OpenAI released GPT-5.3-Codex, which combines the cutting-edge coding performance of GPT-5.2-Codex with GPT-5.2’s reasoning and professional knowledge capabilities, achieving a 25% speed increase.

MiniMax’s release of M2.5 aims to set a price benchmark among AI programming models of similar performance levels, promoting significant development in agent applications.

For example, MiniMax M2.5 adopts Agent RL (Reinforcement Learning) algorithms and Reward design.

At the algorithm level, MiniMax M2.5 continues to use the CISPO algorithm proposed by its team, ensuring the stability of MoE (Mixture of Experts) models during large-scale training.

The CISPO algorithm, developed by the MiniMax team, is a reinforcement learning algorithm mainly used to optimize the training process of large language models.

To address the credit assignment problem caused by long context lengths in agent scenarios, MiniMax M2.5 introduces a process reward mechanism to monitor the quality of generation across the entire chain.

Additionally, to deeply align with user experience, MiniMax M2.5 evaluates task completion time through agent trajectory assessment, achieving an optimal balance between model intelligence and response speed.

Public information shows that MiniMax has grown into a leading global AI multimodal company in just four years, with total investment of about 500 million USD, compared to OpenAI’s cumulative expenditure of approximately 40 to 55 billion USD.

(Source: China Fund News)