The Chinese AI company, MiniMaxAI, has just launched a large-scale open-source reasoning model, named MiniMax-M1. The model, released on Day 1 of the 5-day MiniMaxWeek event, seems to give a good competition to OpenAI o3, Claude 4, DeepSeke-R1, and other contemporaries. Along with the chatbot, MiniMax has also released an agent in beta version, capable of running code, building apps, creating presentations, and more. In this article, we\u2019ll explore the key features of MiniMax-M1, learn how to access it, and test it out on a few tasks. We\u2019ll also be exploring the MiniMax Agent, so read till the end to watch the agent in action!<\/span><\/p>\n MiniMax\u2011M1 is an open\u2011source, large\u2011scale, hybrid\u2011attention reasoning model, developed by Shanghai\u2011based AI startup MiniMax. <\/span>The thinking model comes with a web search feature and can handle multimodal input in the form of text, images, presentations, and more across various formats.<\/p>\n Built on a Mixture\u2011of\u2011Experts (MoE) architecture<\/a>, the model is trained on a total of 456\u202fbillion parameters, with about 45.9\u202fbillion activated per token. Moreover, the model is released under an Apache\u202f2.0 license, making it truly open-source.<\/span><\/p>\n MiniMax has introduced Lightning Attention for its M1 model, dramatically reducing inference costs. To put it in numbers, it uses just 25% of the FLOPs compared to DeepSeek\u2011R1 at 100,000\u2011token generation.<\/span> The model is trained via large\u2011scale RL using CISPO (Clipped Importance Sampling Policy Optimization), which clips sampling weights instead of updates. This led to efficient training on 512 A800 GPUs over 3 weeks, costing only around $534,700. This is far lower than the millions spent by competitors like OpenAI and Google.<\/span><\/p>\n Here are the key features of the new MiniMax-M1 model:<\/span><\/p>\n Also Read: New Deepseek R1-0528 Update is INSANE<\/a><\/em><\/p>\n Here\u2019s how the M1\u201180k model stacks up across major benchmark domains:<\/span><\/p>\n Also Read: What are LLM Benchmarks?<\/a><\/em><\/p>\n MiniMax\u2011M1 is fully open\u2011sourced and available on both GitHub<\/a> and Hugging Face<\/a>. You can also use the model directly on its chatbot interface: <\/span>https:\/\/chat.minimax.io\/<\/span><\/a>.<\/span><\/p>\n The MiniMax API offers structured function-calling interfaces and other chatbot APIs as well. It also provides tools for search, image\/video generation, voice synthesis, and voice cloning \u2013 now tailored for agentic workflows.<\/span><\/p>\n Now that we know how to access MiniMax-M1, let\u2019s try it out. In this section, I\u2019ll be testing out three different prompts to gauge the performance of the model in:<\/span><\/p>\n Let\u2019s get started!<\/span><\/p>\n Prompt: <\/strong>\u201cGenerate a simulation of a red pentagon, rotating clockwise inside a black hexagon. There are 2 balls moving inside the pentagon \u2013 one blue, one yellow \u2013 bouncing off of the walls following the laws of Physics. Every time a ball hits an edge, its colour must change. And every time they collide with each other, they must both become the mixture of both their colours.\u201d<\/span><\/i><\/p>\n \nWhat is MiniMax-M1?<\/h2>\n
![\"\"](\"https:\/\/cdn.analyticsvidhya.com\/wp-content\/uploads\/2025\/06\/MiniMax-input.webp\")
<\/figure>\nKey Features<\/span><\/h3>\n
\n
MiniMax\u2011M1: Benchmark Performance<\/span><\/h2>\n
![\"MiniMax\u2011M1](\"https:\/\/cdn.analyticsvidhya.com\/wp-content\/uploads\/2025\/06\/MiniMax-Benchmark.webp\")
<\/figure>\n\n
How to Access MiniMax\u2011M1<\/span><\/h2>\n
MiniMax-M1: Hands-on Testing<\/span><\/h2>\n
\n
Task 1: Animated Simulation<\/span><\/h3>\n
Output<\/h4>\n