MiniMax-01

Σχετικά με MiniMax

Ιδρύθηκε 2021 · Shanghai, China

MiniMax (上海稀宇科技, Shanghai Xiyu Technology) is a Shanghai-based AI startup founded in late 2021 by Yan Junjie (former vice president of SenseTime). The company has built a portfolio of consumer AI products including the Talkie and Hailuo AI chatbots, abab text models, and the Hailuo Video generative video line. MiniMax has raised over $850M from investors including Alibaba, Tencent, Hillhouse Capital and IDG Capital, with a 2024 valuation above $2.5B. MiniMax-01, released January 2025, was the company's first open-weight frontier release and introduced lightning attention at scale — a linear-attention variant — with a 4M token training context and 1M context inference window.

Αρχιτεκτονική

Hybrid Lightning-Attention + Softmax-Attention Mixture-of-Experts

MiniMax-01 (MiniMax-Text-01 base plus MiniMax-VL-01 vision variant) combines lightning attention — MiniMax's linear-attention design — with periodic softmax attention layers (every 8th layer is full softmax). The architecture has 80 layers with 6,144 hidden size, and MoE feed-forwards with 32 experts and top-2 routing yielding 45.9B active out of 456B total. The hybrid design enables 4M-token effective training context and 1M-token inference context, with near-linear compute per token in the lightning-attention layers. MiniMax describes MiniMax-01 in its technical paper as the first production-scale linear-attention LLM. The VL variant adds 336x336 image patch encoding for vision input. Released January 2025 under the MiniMax Model License with open weights on Hugging Face alongside hosted API access via the MiniMax Open Platform.

Παράμετροι

456B total, 45.9B active per token (32 experts, top-2 routing)

Περιεχόμενο

4.000.000 tokens

Δυνατότητες

• Industry-leading 1M-4M token context window
• First production-scale lightning (linear) attention LLM
• 456B total / 45.9B active parameters
• Strong needle-in-haystack performance reported in the technical paper
• Competitive with GPT-4o on standard benchmarks (MMLU, GSM8K, HumanEval)
• Open weights released — first frontier-scale lightning-attention model in the open
• Vision variant (MiniMax-VL-01) supports image inputs
• Multilingual with strong Chinese and English performance
• Best for: ultra-long-context analysis, Chinese-language applications, long-form agentic workflows, research on linear attention.

Εκπαίδευση & άδεια

Pretrained on trillions of tokens of multilingual web data with heavy Chinese and English representation, code and math; the VL variant adds image-text pairs. Exact data composition is partially described in the technical paper. Knowledge cutoff approximately late 2024.

Άδεια: MiniMax Model License. Permits commercial use with acceptable-use restrictions; products at scale may require registration with MiniMax. Review the license file on Hugging Face before deployment.

Δοκιμές ασφάλειας: MiniMax conducted internal safety evaluation. The model filters politically sensitive topics consistent with Chinese regulations. No public third-party red-team report.