Real-world testing of MiniMax-M1: Understanding why M1 is among the top open-source models globally.

In late June, LMARaena released a statement.Large ModelText-based competitive leaderboard.

The MiniMax-M1 ranked 12th, tied with the DeepSeek V3 0324, DeepSeek R1, and Qwen3. Thus, in China...LLMYu SanjiaOfficially joined forces globallyopen sourceThe first tier of models.

LMArena rankingsThese are real, hard-earned results, blindly selected by users worldwide. They don't look at the parameters; they rely on the "hard skills" of the model itself.

This weekup to dateAccording to the ranking data, the MiniMax-M1 has entered the top 10.

The MiniMax-M1 employs a hybrid expert (MoE) architecture, combined with a lightning-fast attention mechanism, ensuring that only 45.9 B of the 456 billion parameters are actually activated.—Only wake up "when reasoning"Minority"expert.

For the training phase, MiniMax developed its own CISPO, with an official training cost of $535,000, far lower than the previous costs that often started at tens of millions of dollars.Large ModelTraining costs.

Combined with the "minority" activation method, the inference cost of a long response with 100K tokens is only one-quarter of that of DeepSeek-R1.

The MiniMax-M1's mathematical capabilities are also quite impressive. AIAt ME 2024, the MiniMax-M1-80K achieved a score of 86%, placing it directly in the top tier among publicly weighted models.

In LMArena up to dateofLarge ModelIt also consistently ranks first in the math category of the rankings.

MiniMax-M1's contextual capabilities are also...powerfulOutrageous: Input up to 1 million tokens, output 80,000 tokens. You could swallow the entire Water Margin in one go without choking.

While the MiniMax-M1's code generation capabilities aren't top-tier, they are still relevant in LMArena. Programming RankingIt remains firmly in the second tier, demonstrating excellent performance in instruction alignment, logical clarity, and code readability.

adaptation Agent There is absolutely no pressure in terms of scenarios and auxiliary development tasks.

But talk is cheap; let's do some real testing and see how powerful it really is.

General scenarios

Prompt wordsPlease plan a week-long trip for me, starting from Beijing and ending in Xinjiang, with a budget of under 10,000 yuan. Please also consider the weather in July and provide transportation options.

The transportation, accommodation, and sightseeing expenses were all well planned, and the time allotted for visiting each attraction was also well planned. Compared to other travel guides on Xiaohongshu (a Chinese social media platform), this plan is truly superior.

math

Prompt wordsFind the number of integers n < 1000 that can be written as 2.^a-2^b The number of positive integers (a>b≥0).

This is 2021 AIQuestion 3 in the ME test tests enumeration, quadratic judgment, and...fastThe upper bound estimate suggests that being able to answer the question smoothly indicates a solid foundation in basic algebra.

Web development scenarios

Prompt wordsCreate a set of 3D cards in a 3x3 grid that flips and slightly enlarges when hovered over the screen. Use only HTML and CSS (no JS required). Content is displayed correctly on both sides. The flipping process is smooth and realistic.

Creating a flip animation using pure HTML+CSS may seem like a simple trick at first glance, but it actually tests the underlying structure—whether the three chains of layout, 3D geometry, and micro-interactions can be seamlessly integrated.

MiniMax M1Prompt wordsThe design was followed very well, and the slightly enlarged effect after the card was flipped was also presented very well.

Prompt wordsCreate a sticky note wall where users can create, edit, drag, and delete colored sticky notes.

A colorful sticky note wall may look like a toy, but it pulls event models, state persistence, drag-and-drop geometry, and interface scheduling all into the red zone. Only by running them smoothly can you truly master them.

The MiniMax-M1's success is not a fluke, but a concentrated reflection of MiniMax's long-term strategy. Of course, MiniMax has more than one trump card in its hand:

Hailuo 02, a video generation model, has climbed to second place on major global leaderboards.The film excels in complex action sequences, camera work, and realistic lighting, achieving a level of realism approaching that of a film.

In terms of audio models, MiniMax has also achieved several world firsts.From TTS to voice style transfer, its fidelity, rhythm, and naturalness are all top-notch in the industry.

More importantly, MiniMax doesn't follow the old path of brute force to achieve miracles. They don't blindly believe in parameters, but instead focus their efforts on "inference ability" and "engineering steady state"—using fewer resources to run more stably, more accurately, and faster.

MiniMax doesn't intend to launch a single attack; instead, it's pushing language, video, and audio into the open arena simultaneously.

The leaderboard reflects both user votes and market decisions. Establishing a foothold in the open rankings signifies market acceptance.

China AI Initially, it was all about computing power and funding; now it's about ecosystem and speed of implementation. MiniMax uses...open sourceBy employing three key strategies—long context, cross-modal approach, and strategic thinking—China has placed itself in a global arena of competition and incorporated the "Chinese approach" into the rules of the game.

Going forward, no matter how the capital winds shift, as long as MiniMax can maintain this positive cycle of technology, products, and ecosystem, it will remain at the table.