How MiniMax’s Open‑Source Linear‑Attention Model Is Shaking Up the Global AI Landscape

MiniMax-01 Series Open‑Source Release

In January 2024 MiniMax released the MiniMax‑01 series under an open‑source license. The models implement a large‑scale linear attention mechanism, reducing attention complexity from quadratic to linear while preserving accuracy comparable to leading industry large language models.

Model Portfolio

Trillion‑parameter expert mixture‑of‑experts (MoE) : Scales beyond 1 trillion parameters using sparse routing to keep inference cost low.

Low‑latency speech models : Optimized for major Asian languages, achieving sub‑100 ms response times on typical server‑grade GPUs.

Multimodal capabilities : Unified architecture supports text, speech, music, image, and video generation.

Consumer Applications

Talkie : An AI chat application that enables users to converse with virtual characters. Launched internationally in June 2023; by June 2024 it recorded over 11 million monthly active users in the United States and ranked among the top five free entertainment apps.

Hailuo AI platform : Provides text and music generation services. In September 2024 the platform introduced video‑01, a text‑to‑video model whose benchmark performance is reported to be on par with OpenAI’s Sora model.

Enterprise API Platform

MiniMax offers a secure, flexible API that allows developers and enterprises to integrate multimodal LLM capabilities into their products. The service emphasizes low latency, role‑based access control, and scalable deployment on both on‑premise and cloud environments.

International Expansion Strategy

To mitigate domestic regulatory constraints, MiniMax has established overseas entities and operates on foreign servers, enabling access to global markets. Ongoing geopolitical tensions between the United States and China may impact this strategy.