MiniMax Text 01 (MiniMax)

MiniMax Text 01 is a large-scale conversational model released by the Chinese AI lab MiniMax AI, the Shanghai-based company founded in late 2021 by former SenseTime researchers. It belongs to the MiniMax-01 family of open models, which were among the first to combine a Lightning Attention mechanism with a Mixture-of-Experts (MoE) architecture. This hybrid design was specifically created to handle ultra-long contexts, allowing training with sequences up to the cataloged API limit and inference-time extension to roughly four the cataloged API limit, a capability that set MiniMax AI apart in long-context research at the time of release. Within that family, Text-01 sits alongside the multimodal sibling VL-01 as the language-focused variant, with both published under the permissive MIT license to encourage downstream experimentation and self-hosting.

With roughly 456 billion total parameters distributed across its MoE experts, MiniMax Text 01 is engineered for assistant and agent workloads that demand very large working context and predictable per-token economics, making it most useful for document-heavy retrieval, multi-file code reasoning, long conversation summarization, and pipeline-style agent tasks that benefit from holding an entire corpus in context rather than chunking it. In practice, the open-weight release and the Lightning Attention design mean that teams with sufficient hardware can self-deploy the model rather than relying on a hosted endpoint, which fits well for private enterprise knowledge bases, research workflows operating over entire codebases or book-length documents, and developer tooling where both extreme context length and fine-grained control over serving are priorities.