MiniMax M3 Achieves Three Key Features Simultaneously with Open Source Technology

Chinese AI company MiniMax has unveiled its open-source large language model, “M3,” which is making waves in practical benchmark tests. Built with approximately 428 billion parameters, the AI security-specialized agent “MopMonk,” based on this model, achieved a 73.1% success rate on the cybersecurity evaluation benchmark “CyberGym.” This score rivals OpenAI’s models and places it at the forefront of Chinese AI technologies.

Developed by the University of California, Berkeley (UC Berkeley), CyberGym is a trusted industry benchmark that evaluates real-world scenarios like jailbreak attacks and data leakage countermeasures. It uses 1,507 real-world vulnerabilities accumulated from Google’s OSS-Fuzz project as test cases.

What stands out is that while competitors such as the GPT series have reached a parameter size of 1 trillion, M3 delivers comparable performance with less than half the parameters. This makes M3 the first open-source model to achieve three cutting-edge features—1M (approximately 1 million tokens) ultra-long context, native multimodal processing, and advanced code generation—simultaneously at a top-tier level. While several models excel in individual metrics, only flagship closed-source models like GPT or Claude have been able to meet all three criteria at once.

China’s tech media outlet “TMTPost” conducted empirical tests on M3’s multimodal and code execution capabilities. Below, we analyze M3’s capabilities based on these findings.

Testing Multimodal Capabilities

In enterprise AI agents, multimodal processing capabilities are crucial. When integrated into real-world business workflows, these agents are expected to handle long contexts, understand multimodal content like images and videos, and execute tasks via code. M3 leverages MiniMax’s expertise in video generation foundational models, integrating visual comprehension natively into its language model.

The empirical tests included a two-stage video comprehension task. In the first stage, technology-related videos from the Chinese platform Bilibili were used as input, and the AI was tasked with providing a summary in under 300 characters, a core conclusion, and three supporting arguments with corresponding timestamps. Notably, the AI needed to understand not just subtitles but also the relationship between graphs or data displayed on the screen.

M3 not only generated conclusions aligned with the video content but also accurately identified three key timestamps and linked each supporting argument to specific moments in the video. This demonstrated its capability to deconstruct long-form videos into traceable chains of evidence.

In the second stage, an English programming tutorial video was used. The task required the AI to break down the instructions into Step 1 through Step N, translate core technical terms, timestamp each step, and finally produce a structured Chinese tutorial document. M3 delivered a comprehensive document including data preparation, goal definition, implementation methods, comparative analysis, application examples, anti-pattern warnings, and a summary, with each line tagged with precise time ranges.

Code Generation and Visualization Capabilities

M3’s combination of multimodal understanding and code generation opens the door to highly practical applications. In one test, video analysis results were converted into a complete, executable single-file HTML page. The requirements for the HTML included a fixed top navigation bar, a conclusion card in the first view, an area for displaying supporting arguments, placeholder images for screenshots linked to timestamps, dark mode compatibility, and responsive design for both desktop and tablet devices.

The generated HTML included a title area, a conclusion card, indicator cards, and a dark-styled page with time anchors. The indicator cards displayed performance metrics, and the interface featured thoughtful design elements like “return to video scene at three time anchors to verify evidence.” A single video was effectively transformed into a navigable, clickable information page.

To further test the model, the same video was used for more complex tasks, such as extracting quantitative test data (performance, power consumption, frame rates, etc.) from visuals and subtitles. This data was structured into JSON format and used to generate an interactive data dashboard based on ECharts, including four core indicator summary cards, a dropdown filter for test categories, bar graphs, and trend comparison graphs. M3 seamlessly handled data extraction, formatting specifications, frontend implementation, and interaction logic.

Implications for the Industry

The emergence of M3 demonstrates that the gap between open-source and closed-source models is rapidly narrowing. Achieving the simultaneous milestones of multimodality, ultra-long context, and code generation is a significant milestone for enterprise AI agents.

In the past, discussions around AI model performance often revolved around issues like the so-called “AI Tokenpocalypse” cost problem or inefficient solutions like reducing token usage. With M3, which achieves high inference efficiency despite a smaller parameter size, a practical solution to cost concerns has emerged.

However, it is worth noting that the current results come from Chinese media tests, and third-party verification is still awaited. Additionally, it is important to confirm whether the videos used for evaluation were representative and culturally unbiased.

As highlighted in the context of GLM 5.2 and the Shrinking Profit Margins in AI Inference, the competition in AI model performance is intensifying. The rise of open-source models like M3 could significantly impact the revenue structures of the AI industry.

Editorial Opinion

In the short term, M3’s performance could lower the barriers to adopting AI agents. For companies seeking to move away from closed-source API dependencies, the emergence of an open-source option that combines multimodality and coding is welcome news. However, further validation is needed regarding operational stability, latency, and multilingual support quality, including Japanese.

From a long-term perspective, the improved performance of open-source models cannot be overlooked in its potential impact on the business models of closed-source giants. Should models like M3, which achieve “three key features simultaneously,” become widespread, the scope of AI agents could expand beyond text-centric use cases to include multimodal tasks involving video, audio, and images. As suggested by the First Autonomous Ransomware Attack by an AI Agent, advances in agent capabilities could also introduce new security risks. This is a critical question our editorial team would like to raise.

References

• TMTPost: “MiniMax M3 Breaks Records: Challenging Silicon Valley’s Closed-Source Giants?” — Published on July 8, 2026