Join the event entrusted by business leaders for almost two decades. VB Transf transform gathers the people who build a real strategy of AI. Get more information
Beginning of the Chinese AI Minimax, perhaps better known in the West by its realistic success Oh Hailuo video modelhas launched their latest model of great language, Minimax-m1 – And with great news for companies and developers, it’s completely Source Open under an Apache 2.0 licenseThat is, companies can take it and use it for commercial applications and modify them to their taste without restrictions or payments.
M1 is an open weight offer that establishes new standards in long context reasoning, use of agents tools and efficient calculation performance. Today is available in the AI code sharing community Hugging face and Microsoft’s rival code shares the GitHub communityThe first launch of what the company was called as “minimaxweek” of its social account in X, with more expected products ads.
MINIMAX-M1 is distinguished with a context window of one million entry sheets and up to 80,000 output tiles, positioning it as one of the most expansive models available for long context reasoning tasks.
The “context window” in models of great language (LLMS) refers to the maximum number of tiles that the model can process at the same time, including entry and output. The sheets are the basic text units, which can include whole words, word parts, punctuation marks or code symbols. These sheets become numerical vectors that the model uses to represent and manipulate the meaning through its parameters (weights and biases). They are, in essence, the mother tongue of the LLM.
For comparison, OPENAI’s GPT-4O Has a context window of only 128,000 tiles, enough to change about information about a novel between the user and the model in a single interaction back and forward. At 1 million tiles, Minimax-M1 could exchange a small collection or information in the book series. Google Gemini 2.5 Pro offers a higher limit of 1 -million token contextIn addition, with a window of 2 million works in the works.
But M1 has another trick on his sleeve: it has been formed through reinforcement learning in an innovative, resource and very efficient technique. The model is formed by a hybrid mixture (MOE) with a ray attention mechanism designed to reduce inference costs.
According to the technical report, Minimax-M1 consumes only 25% of floating points (flops) required by DeepSeek R1 to a generation of 100,000 tiles.
Architecture and variants
The model is presented in two variants: Minimax-M1-40K and Minimax-M1-80K, which are collected in their “thought budgets” or output length.
The architecture is based on the previous Foundation of Minimax-Text-01 of the company and includes 456 billion parameters, with 45.9 billion activities per witness.
A prominent feature of launch is the model’s training cost. MINIMAX reports that the M1 model was formed through large -scale reinforcement learning (RL) in an efficiency rarely seen in this domain, with a total cost of $ 534,700.
This efficiency is accredited in a personalized RL algorithm called CISPO, which clips the importance of weights sampling instead of Token updates and the hybrid care design that helps streamline the scale.
This is a surprisingly “cheap” amount for a Frontier Llm, as Deepseek trained their R1 reasoning model A costs of 5 to 6 million dollarsWhile the Training Cost of Openis’ GPT-4, a model of more than two years, was now Is said to exceed $ 100 million. This cost comes from the price of the Graphic Processing Units (GPU), the massively parallel computer hardware mainly manufactured by companies such as NVIDIA, which can cost between $ 20,000 and $ 30,000 per module, and the energy needed to run these chips continuously in large -scale data centers.
Reference performance
MINIMAX-M1 has been evaluated through a series of established reference points that test advanced reasoning, software engineering and tool use capabilities.
In Aime 2024, a reference point for math competition, the M1-80K model marks 86.0% accuracy. It also offers a strong performance in coding and long context tasks, achieving:
- 65.0% in LivecodeBench
- 56.0% in Swe-Cench
- 62.8% in Tau-Cench
- 73.4% to Openai MRCR (version 4-B
These results place Minimax-M1 ahead of other open-weight competitors such as Deepseek-R1 and Qwen3-23b-A22b In various complex tasks.
While closed weight models such as Openai’s O3 and Gemini 2.5 Pro are still some reference points, Minimax-M1 restricts the performance gap considerably, while keeping free accessible under an Apache-2.0 license.
For the deployment, Minimax recommends VLLM as a Backend of Service, citing its optimization for workloads of large models, memory efficiency and handling of lot applications. The company also provides deployment options through the Transformers Library.
MINIMAX-M1 Includes calls for structured functions call and packaged with a Chatbot API with online search, video and image generation, speech synthesis and voice cloning tools. These features aim to support wider agents in actual world applications.
Implications for those responsible for technical decisions and business buyers
The open access of Minimax-M1, the capacities of the long context and the calculation of efficiency address several recurring challenges for the technical professionals responsible for managing the AI systems on a scale.
For engineering, those responsible for the full life cycle of the LLM, such as optimizing model performance and deployment in tight terms, Minimax-M1 offers a lower operating cost profile while supporting advanced reasoning tasks. Its long context window could significantly reduce pre -procedural efforts for business documents or registration data that cover tens or hundreds of thousands of sheets.
For those who manage AI orchestration pipes, the ability to refine and deploy minimax-m1 through established tools such as VLLM or transformers supports an easier integration into the existing infrastructure. Hybrid care architecture can help to simplify the model of climbing and competitive performance of the model at the Reference Points of Reasoning and Software Engineering in various steps offers a high capacity base for internal co -pilots or agents based systems.
From the perspective of the data platform, the equipment responsible for maintaining an efficient and scalable infrastructure can benefit from the support of M1 for calls for structured functions and their compatibility with automated pipes. Its open source nature allows equipment to adapt the performance to their stack without vendor blocking.
Safety advantages may also find value in the M1 potential assessment to deploy a M1 high capacity capacity model in the transmission of sensitive data to the end of third parties.
As a whole, Minimax-M1 presents a flexible option for organizations that seek to experiment or expand the advanced AI capabilities while managing costs, staying within operating limits and avoiding ownership restrictions.
The launch indicates the continuous focus of Minimax on practical and scalable models. Combining open access with advanced architecture and calculation efficiency, Minimax-M1 can serve as a founding model for developers who build new generation applications that require depth of reasoning and long-range entry understanding.
We will continue to follow other minimax throws throughout the week. Stay tuned!
Source link
