MiniMax: MiniMax-01

MiniMax-01 is presented as a composite endpoint that routes text generation through MiniMax-Text-01 and image understanding through MiniMax-VL-01, allowing a single request to mix written prompts with visual inputs while producing text-only outputs. The architecture inherits the sparse-expert design of its text foundation, which combines 456 billion total parameters with 45.9 billion activated per inference, a pattern that keeps per-request compute modest even at the scale of a frontier-sized model. Because the visual branch is delegated to a companion VL model rather than fused into every layer, MiniMax-01 is best understood as a coordinated pair of specialists rather than a monolithic multimodal network.

From a practical standpoint, the model targets workloads that pair documents or screenshots with long written context, supported by a million-token window suitable for large transcripts, code repositories, or multi-document retrieval tasks. The underlying text backbone has been benchmarked competitively with leading contemporaries on academic suites and reportedly achieves full recall on needle-in-a-haystack tests at very long context lengths, which suggests that summarization, grounded question answering, and extended reasoning chains are realistic deployment scenarios. Creators also describe the hybrid lightning-attention design as a way to keep long-input costs closer to linear, making MiniMax-01 a fit for high-volume, context-heavy pipelines where predictable latency and pricing matter as much as raw capability.