Superseded (verified 2026-09-30)

Text-01 (early 2025) predates M1, M2.x and minimax-m3. Keep for history.

MiniMax-Text-01

MiniMax’s first large open-weight language model.

Specifications (Hugging Face card)

• Parameters: 456B total / 45.9B active per token; 80 layers
• Architecture: hybrid of Lightning Attention, softmax attention and MoE
• Context: trained to 1M tokens; up to 4M tokens at inference (the earlier note said only 1M)
• Licence: custom Model Agreement for the weights, MIT for code
• Reported scores: MMLU 88.5; LongBench v2 56.5 (vendor-reported)
• Release date: the card text fetched gives none; the published date is the commonly reported January 2025, unverified here.

See also

• MiniMax-VL-01, MiniMax M1, minimax-m3

Sources

• https://huggingface.co/MiniMaxAI/MiniMax-Text-01 (accessed 2026-09-30)