At the end of June, LMArena released a large model text arena leaderboard.
MiniMax-M1 It ranked 12th, tied with DeepSeek V3 0324, DeepSeek R1, and Qwen3. At this point—domestic LLMsthe Big Threeofficially joined forces in the global top tier of open-source models.
The LMArena leaderboard,is a real, hard-earned record of global users' blind votes. It doesn't look at parameters; it's a contest of the model's own "hard skills."
The latest leaderboard data this week shows that MiniMax-M1 has already entered the top 10.
01. Why MiniMax-M1 Ranks Among the Top
MiniMax-M1 adopts a Mixture of Experts (MoE) architecture, combined with a lightning attention mechanism. Of its 456 billion parameters, only 45.9B are actually activated—at inference, it only wakes up the "minority" experts.
For training, MiniMax developed CISPO in-house. The officially announced training bill was $535,000, far below the large model training costs that once started at tens of millions of dollars.
Combined with the "minority" activation method, the inference cost of a 100K-token long response is only one quarter that of DeepSeek-R1.
MiniMax-M1's math ability is also no less impressive. On AIME 2024, MiniMax-M1-80K achieved a score rate as high as 86%, directly reaching the top tier among openly weighted models;
In LMArena's latest large model leaderboard—math category, it also firmly holds first place.
MiniMax-M1's context capability is likewise absurdly powerful: input up to 1 million tokens and output 80,000 tokens. It can swallow the entire "Water Margin" in one go without choking.
Although MiniMax-M1's code generation ability is not in the top tier, on LMArena the programming leaderboardit firmly holds a place in the second tier, performing excellently in instruction alignment, logical clarity, and code readability.
Adapting to Agent scenarios and assisting development tasks is no pressure at all.
02. Hands-On with MiniMax-M1
But talking more is not as good as doing more. Let's test it together and see how strong it really is.
General Scenarios
Prompt: Plan a one-week travel itinerary for me, departing from Beijing, destination Xinjiang, budget within 10,000 yuan, considering July weather and providing transportation options.
Transportation and accommodation, as well as attraction expenses, were all planned very well, and the visit duration for each attraction was also well planned. Compared with the guides on Little Red Book, the entire plan is truly even better.
Mathematics
Prompt: Find the number of positive integers n<1000 that can be written as 2a−2b (a>b≥0).
This is the 3rd problem in the 2021 AIME test. It examines enumeration, quadratic judgment, and fast upper-bound estimation. Being able to solve it smoothly shows that basic algebra is equally solid.
Web Development Scenario
Prompt: Create a set of nine-grid 3D card components that flip on hover and slightly enlarge. Use only HTML and CSS (no JS). Both the front and back sides have content and display normally. The flipping process is smooth and realistic.
Using pure HTML+CSS for flip animations may seem like a small trick at first glance, but it actually tests the fundamentals—whether the three chains of layout, 3D geometry, and micro-interactions can mesh in one go.
MiniMax M1 follows the prompt very well, and the effect of the card slightly enlarging after flipping is also presented nicely.
Prompt: Create a sticky note wall where users can create, edit, drag, and delete colorful sticky notes.
A colorful sticky-note wall looks like a toy, yet it pulls the event model, state persistence, drag geometry, and interface scheduling all into the red zone; running it smoothly is the real test of skill.
03. Minimax Technology Panorama
MiniMax-M1's highlight is not a fluke, but a concentrated expression of MiniMax's long-term strategy. Of course, MiniMax has more than just this one ace up its sleeve:
The video generation model Hailuo 02 has climbed to second place on major global leaderboards. Its complex motion, camera scheduling, and photorealistic lighting and shadow capabilities are all strong, with a sense of realism approaching cinematic quality.
In audio models, MiniMax has also taken multiple global first places.From TTS to voice style transfer, its fidelity, rhythm, and naturalness are top-tier in the industry.
More critically, MiniMax is not taking the old path of brute force producing miracles. They do not fetishize parameters, but instead spend their energy on "reasoning ability" and "engineering stability"—using fewer resources to run more stably, more accurately, and faster.
04. Some Sharing
MiniMax did not intend a single-point breakthrough, but instead pushed language, video, and audio onto the public arena at the same time.
On the leaderboards, users are voting, and the market is voting too. Once it can hold its ground in open rankings, it represents market recognition.
Chinese AI initially competed on computing power and funding, but now it is competing on ecosystem and speed of implementation; MiniMax, with its three cards of open source, long context, and cross-modality, has placed itself in the global arena of competition and has also written the "Chinese playbook" into the rules of the game.
Next, no matter how the capital winds shift, as long as MiniMax can maintain this positive cycle of technology—product—ecosystem, it will keep sitting at the table.
Original link:Breaking into the first tier with a fraction of a big company's budget, what strength does MiniMax actually have?