MiniMax-Text-01

MiniMax-Text-01 is the text flagship of MiniMax's 01 series, introduced with a technical paper in January 2025. It is a 456-billion-parameter mixture-of-experts model (45.9B active per token, 32 experts) built around 'lightning attention' — a linear-attention mechanism combined with softmax attention and MoE routing. The model trains on a 1-million-token context and extrapolates to 4 million tokens at inference time, at which point it was the longest-context open-weight model available. MiniMax reported GPT-4o/Claude-3.5-class scores on standard benchmarks (75.7 MMLU-Pro, 54.4 GPQA) with 20–32x the context window of those systems, and released the weights openly alongside a sibling vision-language model, MiniMax-VL-01.\n\nThe honest picture: this is a genuine open-weights giant, but 456B total parameters means practical deployment needs serious infrastructure — the API is the realistic route for most teams. Benchmark numbers are MiniMax's own, and the 'open' release predates MiniMax's later shift toward agentic consumer products. Check the release license terms before commercial use.