MiniMax M3 Sets New Standard for Long-Context Enterprise AI
MiniMax has officially launched its latest artificial intelligence system, the M3 Multimodal Foundation Model, introducing a powerful new platform capable of processing up to one million tokens in a single context window. Announced on June 1, 2026, the model supports text, image, and video inputs, making it suitable for complex enterprise workflows and large-scale data analysis.
The release marks a significant advancement in long-context AI processing, enabling organizations to analyze extensive datasets, lengthy documents, large code repositories, and multi-document workflows more efficiently than previous AI generations.
Advanced Multimodal Capabilities
The MiniMax M3 model is designed to handle multiple data formats simultaneously, including text, images, and videos. Developers can access the model through the EvoLink platform using OpenAI-compatible API endpoints, simplifying integration into existing applications and enterprise systems.
This multimodal approach allows businesses to automate document understanding, spreadsheet analysis, enterprise agents, and knowledge management tasks while maintaining context across exceptionally large datasets.
Innovative Sparse Attention Architecture
At the core of the new model is MiniMax Sparse Attention, a proprietary architecture developed to efficiently process ultra-long contexts. Unlike traditional compression techniques, the system operates on a standard Grouped-Query Attention backbone while using block-level selection on real, uncompressed Key-Value memory blocks.
This design significantly reduces computational requirements without relying on aggressive data compression techniques that may impact performance.
Significant Performance Improvements
According to MiniMax, the M3 model delivers substantial speed improvements over its predecessor, M2. The company reports decoding speeds that are 15.6 times faster and prefill processing speeds that are 9.7 times faster when working with contexts approaching one million tokens.
The model also supports a maximum output length of 131,072 tokens per request, providing the ability to generate extensive reports, analyses, and structured outputs from large-scale inputs.
Focus on Reasoning Quality
To maintain reasoning quality while reducing computational costs, MiniMax developed a new sub-quadratic framework. This approach aims to overcome limitations that have affected previous long-context AI systems. However, while MiniMax has published performance benchmarks related to speed and efficiency, independent third-party evaluations of reasoning quality and accuracy have not yet been released.
Market Position and Enterprise Strategy
Founded in 2022, MiniMax has rapidly expanded its presence in the AI sector. The company became publicly listed on the Hong Kong Stock Exchange in January 2026 and is backed by major technology investors including Tencent, Alibaba, and miHoYo. MiniMax currently serves more than 200 million users through its consumer-facing applications.
The M3 API is priced at $0.30 per million input tokens and $1.20 per million output tokens, positioning it as a cost-effective option for organizations requiring large-context AI processing.
Future Impact
With its combination of multimodal processing, ultra-long context handling, and competitive pricing, the MiniMax M3 model could significantly influence enterprise AI adoption. As demand for AI systems capable of understanding complex datasets continues to grow, MiniMax positions itself as a notable competitor in the global AI landscape.
Frequently Asked Questions
What is the MiniMax M3 model?
The MiniMax M3 is a multimodal foundation AI model that supports text, image, and video inputs with a context window of up to one million tokens, designed for enterprise-scale data processing and analysis.
How much faster is M3 compared to M2?
MiniMax reports that M3 delivers decoding speeds 15.6 times faster and prefill processing speeds 9.7 times faster than its predecessor when handling near-maximum context lengths.
What is MiniMax Sparse Attention?
MiniMax Sparse Attention is a proprietary architecture that processes ultra-long contexts efficiently using block-level selection on uncompressed Key-Value memory blocks, reducing computational costs without heavy data compression.
How much does the MiniMax M3 API cost?
The M3 API is priced at $0.30 per million input tokens and $1.20 per million output tokens, making it a competitive option for enterprise organizations.
Who backs MiniMax?
MiniMax is backed by major technology investors including Tencent, Alibaba, and miHoYo. The company listed on the Hong Kong Stock Exchange in January 2026.
