The Chinese language AI firm, MiniMaxAI, has simply launched a large-scale open-source reasoning mannequin, named MiniMax-M1. The mannequin, launched on Day 1 of the 5-day MiniMaxWeek occasion, appears to present a superb competitors to OpenAI o3, Claude 4, DeepSeke-R1, and different contemporaries. Together with the chatbot, MiniMax has additionally launched an agent in beta model, able to operating code, constructing apps, creating displays, and extra. On this article, we’ll discover the important thing options of MiniMax-M1, learn to entry it, and check it out on a couple of duties. We’ll even be exploring the MiniMax Agent, so learn until the tip to observe the agent in motion!
What’s MiniMax-M1?
MiniMax‑M1 is an open‑supply, massive‑scale, hybrid‑consideration reasoning mannequin, developed by Shanghai‑based mostly AI startup MiniMax. The considering mannequin comes with an online search function and might deal with multimodal enter within the type of textual content, photos, displays, and extra throughout numerous codecs.
Constructed on a Combination‑of‑Consultants (MoE) structure, the mannequin is skilled on a complete of 456 billion parameters, with about 45.9 billion activated per token. Furthermore, the mannequin is launched underneath an Apache 2.0 license, making it really open-source.
MiniMax has launched Lightning Consideration for its M1 mannequin, dramatically lowering inference prices. To place it in numbers, it makes use of simply 25% of the FLOPs in comparison with DeepSeek‑R1 at 100,000‑token technology. The mannequin is skilled by way of massive‑scale RL utilizing CISPO (Clipped Significance Sampling Coverage Optimization), which clips sampling weights as an alternative of updates. This led to environment friendly coaching on 512 A800 GPUs over 3 weeks, costing solely round $534,700. That is far decrease than the thousands and thousands spent by opponents like OpenAI and Google.
Key Options
Listed below are the important thing options of the brand new MiniMax-M1 mannequin:
- Hybrid Consideration + MoE Effectivity: M1 combines an MoE structure with Lightning Consideration, reaching as much as 75% compute financial savings in comparison with most different fashions.
- Massive Contect Window: M1’s largest spotlight is its huge context window of 1M enter token tokens and upto 80k output tokens, rivaling Google’s Gemini 2.5 Professional.
- Out there in 40k & 80k Pondering Budgets: MiniMax provides M1‑40k (intermediate) and M1‑80k (superior) variations. The 80k variant performs higher on prolonged reasoning and complicated duties.
- Prolonged Agentic & Lengthy‑Context Reasoning: The fashions benchmark performances present that it excels at long-context benchmarks (OpenAI‑MRCR, LongBench‑v2) and agentic device use (TAU‑bench).
Additionally Learn: New Deepseek R1-0528 Replace is INSANE
MiniMax‑M1: Benchmark Efficiency
Right here’s how the M1‑80k mannequin stacks up throughout main benchmark domains:
- AIME 2024: MiniMax-M1 achieved the very best rating (86.0%), outperforming all proprietary fashions like OpenAI o3, Claude 4 Opus, and Gemini 2.5 Professional.
- LiveCodeBench: MiniMax-M1 scored 65.0% on this coding benchmark, simply barely behind OpenAI o3 and Gemini.
- SWE-bench Verified: On this software program engineering benchmark, MiniMax-M1 achieved 56.0%, clearly outperforming DeepSeek-R1 and Qwen3 by over 10%.
- TAU-bench: In relation to agentic device use, MiniMax-M1 scored 62.8%, beating each DeepSeek and Qwen3. This reveals the mannequin’s sturdy tool-using functionality.
- MRCR: The MRCR benchmark checks the fashions’ abilities in multi-round doc evaluation and long-context reasoning. MiniMax-M1 scored 73.4% on the check, coming near prime business fashions and much forward of different open-source ones.
Additionally Learn: What are LLM Benchmarks?
The right way to Entry MiniMax‑M1
MiniMax‑M1 is totally open‑sourced and accessible on each GitHub and Hugging Face. You may as well use the mannequin straight on its chatbot interface: https://chat.minimax.io/.
The MiniMax API provides structured function-calling interfaces and different chatbot APIs as nicely. It additionally gives instruments for search, picture/video technology, voice synthesis, and voice cloning – now tailor-made for agentic workflows.
MiniMax-M1: Arms-on Testing
Now that we all know the way to entry MiniMax-M1, let’s strive it out. On this part, I’ll be testing out three totally different prompts to gauge the efficiency of the mannequin in:
- Coding an animated simulation
- Looking the net in real-time
- Fixing a logical reasoning puzzle
Let’s get began!
Activity 1: Animated Simulation
Immediate: “Generate a simulation of a purple pentagon, rotating clockwise inside a black hexagon. There are 2 balls shifting contained in the pentagon – one blue, one yellow – bouncing off of the partitions following the legal guidelines of Physics. Each time a ball hits an edge, its color should change. And each time they collide with one another, they have to each grow to be the combination of each their colors.”
Output
Evaluate
In about 10 minutes, MiniMax-M1 generated clear and environment friendly code with correct explanations to create the simulation. It got here out to be visually clear and bodily correct, following the immediate precisely. The balls, nonetheless, get caught on the edges at instances, which I discovered to be an error. Additionally, the balls finally get thrown out of the simulation, which I didn’t count on. However on the entire, I did just like the outcome.
Activity 2: Net Search
Now let’s see how nicely the mannequin can search the net in real-time. On this job, I’m going to ask MiniMax M1 to search out me occasions occurring this week in my metropolis. I’ll be turning on the ‘Search’ mode for this job.
Immediate: “What stand-up comedy reveals are occurring in Gurgaon this week?”
Output
Evaluate
MiniMax M1 virtually instantly searched throughout 30 totally different web sites together with District, All Occasions, and even occasion listings on Fb and Instagram. It gathered all the data and gave me an inventory of all of the stand-up occasions occurring round. The search was performed proper, however the mannequin was unable to learn or detect the timing and venue of the occasions. Nonetheless, it listed out all of the supply web sites on the right-hand aspect, making it simpler to entry the reserving websites.
Activity 3: Logical Puzzle
MiniMax M1 is a reasoning mannequin with considering capabilities. So, I’m going to present it a logical puzzle to resolve on this job. I’ll additionally activate the ‘assume’ mode for this to see the mannequin’s considering course of.
Immediate: “4 persons are crossing a bridge at evening, so all of them want a torch – however they simply have one which solely lasts quarter-hour. Alice can cross in a single minute, Ben in two minutes, Cindy in 5 minutes and Don in eight minutes. Not more than two individuals can cross at a time; and when two cross, they must go on the slower individual’s tempo. How do they get throughout in quarter-hour?”
Output
“The optimum resolution to get all 4 throughout in quarter-hour is as follows:
- Alice and Ben cross collectively (2 minutes).
Time elapsed: 2 minutes
Folks on far aspect: Alice, Ben - Alice returns alone (1 minute).
Time elapsed: 3 minutes
Folks on begin aspect: Alice, Cindy, Don - Cindy and Don cross collectively (8 minutes).
Time elapsed: 11 minutes
Folks on far aspect: Cindy, Don, Ben - Ben returns alone (2 minutes).
Time elapsed: 13 minutes
Folks on begin aspect: Alice, Ben - Alice and Ben cross collectively once more (2 minutes).
Time elapsed: quarter-hour
All 4 on far aspect: Alice, Ben, Cindy, Don
Key technique: Use the 2 quickest people (Alice and Ben) to shuttle the torch effectively, minimizing the time slower people (Cindy and Don) spend crossing collectively. This balances the necessity to transfer slower people whereas preserving torch time.”
Evaluate
MiniMax-M1 took about 2 minutes (125.2 seconds to be exact) to consider the issue and work out the answer. The mannequin’s logical thought course of was on level, resulting in the fitting reply. It didn’t must backtrack any of the ideas, however at some factors, it took a bit longer to assume.
MiniMax Agent in Motion
Now let’s get to the good a part of MiniMax’s new launches – the MiniMax Agent. Though at the moment in beta section, its capabilities and numerous use instances make it fairly promising on the earth of AI brokers. To check out its prowess, I gave it a immediate to construct me an interactive web site. Right here’s how that went.
Immediate: “Construct an interactive web site that includes a digital tour of India.
- It should present the nation map with the states marked, letting the person learn the identify of the state and the respective capital metropolis once they hover the mouse over it.
- It should zoom into the chosen state upon clicking, exhibiting all of the well-known monuments and landmarks as consultant icons.
- Once more, their names and site (identify of metropolis) should be proven when hovered above the icon.
- Upon clicking on an icon, a aspect panel should pop up exhibiting precise photos, the situation, visiting instances, and entrance charge of the landmark.
- The person should have the choice of selecting a number of landmarks, and the app/web site should create a sensible itinerary of visiting all these locations, additionally stating what number of days it could take to cowl all of them. This should additionally consider the commute time based mostly on Google Maps.”
Output
And right here’s the web site it constructed for me:
Evaluate
MiniMax Agent did an ideal job at creating an interactive web site to fulfill my necessities. Though the ultimate web site wasn’t precisely how I had defined in my immediate, I have to say the outcomes are spectacular. The agent discovered the data, added the textual content and pictures, acquired API keys, accessed Google Maps and different apps, constructed the entire system, and even examined it – all by itself. It took about 20 minutes to do the entire thing and even supplied documentation of the method, check outcomes, and all different particulars. A free device doing all of this so nicely, is simply mind-blowing!
You may expertise the total web site right here: https://03w1ujb85t.house.minimax.io/
You may as well check out the agent at no cost by clicking right here. When you enroll utilizing your e mail ID, MiniMax provides you 1,000 free credit to spend on operating the agent.
Conclusion
MiniMax‑M1 represents a serious leap in open-source AI. It’s a first-of-its-kind hybrid-attention MoE mannequin, combining scale and compute effectivity. With an astonishing 1M token context window, this new mannequin is able to long-form reasoning and doc understanding. Regardless of the low coaching prices, it reveals aggressive or superior efficiency throughout customary benchmarks.
The MiniMax Agent can also be fairly spectacular, having the ability to create displays, web sites, and apps by itself. The chatbot interface and reside updates on the aspect give customers the sensation of vibe coding. On the entire, MiniMax‑M1 units a brand new customary in open-source mannequin improvement. Mixing technical sophistication, financial effectivity, and accessibility, it has constructed a robust basis for next-generation AI chatbots and brokers. Because it’s free to make use of for everybody, go forward, strive it out, and tell us within the feedback how you discover it.
Sabreena is a GenAI fanatic and tech editor who’s captivated with documenting the newest developments that form the world. She’s at the moment exploring the world of AI and Information Science because the Supervisor of Content material & Progress at Analytics Vidhya.
Login to proceed studying and luxuriate in expert-curated content material.
