MiniMax-VL-01

MiniMax-VL-01 is the vision-language half of the MiniMax-01 series, open-sourced on January 15, 2025 alongside MiniMax-Text-01. It continues training the 456B-parameter MoE backbone with 512 billion vision-language tokens, pairing a Vision Transformer with the text model to handle both text and images.

The architecture mixes lightning attention with softmax attention to reach a 4-million-token inference context, far beyond competing models at the time, while the weights ship under Apache 2.0 on GitHub and Hugging Face. Vendor benchmarks compared it favorably against GPT-4o and Claude 3.5 Sonnet on multimodal understanding.