88Z AI, 16 June — MiniMax has released MiniMax M3, an open-source native multimodal flagship model. With 428B total parameters and 23B active parameters, it utilizes a proprietary MSA sparse attention architecture to reduce long-context computational costs. As the first open-source model trained with multimodal mixing from step zero, M3 ranked first among open-source models on benchmarks including Artificial Analysis Comprehensive Intelligence Index, GDPval-AA, and Code Arena WebDev, earning praise from industry
At a glance
- MiniMax M3 is the first open-source model with multimodal mixed training from step zero, using a proprietary MSA architecture.
- The model ranked first among open-source models on benchmarks like Artificial Analysis and received praise from industry leaders such as the Vercel CEO.
The update
MiniMax has officially open-sourced its native multimodal flagship model, MiniMax M3. Featuring 428B total parameters and 23B active parameters, the model employs a proprietary MSA sparse attention architecture to efficiently reduce computational costs for long contexts. Notably, M3 is the first open-source model to undergo multimodal mixed training starting from step zero. Following its release, MiniMax M3 achieved top rankings among open-source models on several authoritative benchmarks, including the Artificial Analysis Comprehensive Intelligence Index, GDPval-AA, and Code Arena WebDev. This breakthrough has garnered widespread recognition, receiving positive feedback from prominent industry figures such as the CEO of Vercel and the founder of Y Combinator.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- MiniMax 稀宇科技Source published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.