Chinese AI is making significant strides, challenging leading models like GPT-4, Claude, and Grok with cost-effective, open-source alternatives such as DeepSeek-V3 and Qwen 2.5. These models excel due to their efficiency, accessibility, and strong performance. Many operate under permissive commercial licenses, broadening their appeal to developers and businesses.

MiniMax-Text-01, the newest addition to this group, sets a new standard with its unprecedented 4 million token context length—vastly surpassing the typical 128K-256K token limit. This extended context capability, combined with a Hybrid Attention architecture for efficiency and an open-source, commercially permissive license, fosters innovation without high costs.

Let's delve into MiniMax-Text-01's features:

Table of Contents

• Hybrid Architecture
• Mixture-of-Experts (MoE) Strategy
• Training and Scaling Strategies
• Post-Training Optimization
• Key Innovations
• Core Academic Benchmarks
  • General Tasks Benchmarks
  • Reasoning Tasks Benchmarks
  • Mathematics & Coding Tasks Benchmarks
• Getting Started with MiniMax-Text-01
• Important Links
• Conclusion

Hybrid Architecture

MiniMax-Text-01 cleverly balances efficiency and performance by integrating Lightning Attention, Softmax Attention, and Mixture-of-Experts (MoE).

• 7/8 Linear Attention (Lightning Attention-2): This linear attention mechanism drastically reduces computational complexity from O(n2d) to O(d2n), ideal for long-context processing. It uses SiLU activation for input transformation, matrix operations for attention score calculation, and RMSNorm and sigmoid for normalization and scaling.
• 1/8 Softmax Attention: A traditional attention mechanism, incorporating RoPE (Rotary Position Embedding) on half the attention head dimension, enabling length extrapolation without sacrificing performance.

Mixture-of-Experts (MoE) Strategy

MiniMax-Text-01's unique MoE architecture distinguishes it from models like DeepSeek-V3:

• Token Drop Strategy: Employs an auxiliary loss to maintain balanced token distribution across experts, unlike DeepSeek's dropless approach.
• Global Router: Optimizes token allocation for even workload distribution among expert groups.
• Top-k Routing: Selects the top-2 experts per token (compared to DeepSeek's top-8 1 shared expert).
• Expert Configuration: Utilizes 32 experts (vs. DeepSeek's 256 1 shared), with an expert hidden dimension of 9216 (vs. DeepSeek's 2048). The total activated parameters per layer remain the same as DeepSeek (18,432).

Training and Scaling Strategies

• Training Infrastructure: Leveraged approximately 2000 H100 GPUs, employing advanced parallelism techniques like Expert Tensor Parallelism (ETP) and Linear Attention Sequence Parallelism Plus (LASP ). Optimized for 8-bit quantization for efficient inference on 8x80GB H100 nodes.
• Training Data: Trained on roughly 12 trillion tokens using a WSD-like learning rate schedule. The data comprised a blend of high- and low-quality sources, with global deduplication and 4x repetition for high-quality data.
• Long-Context Training: A three-phased approach: Phase 1 (128k context), Phase 2 (512k context), and Phase 3 (1M context), using linear interpolation to manage distribution shifts during context length scaling.

Post-Training Optimization

• Iterative Fine-Tuning: Cycles of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), using Offline DPO and Online GRPO for alignment.
• Long-Context Fine-Tuning: A phased approach: Short-Context SFT → Long-Context SFT → Short-Context RL → Long-Context RL, crucial for superior long-context performance.

Key Innovations

• DeepNorm: A post-norm architecture enhancing residual connection scaling and training stability.
• Batch Size Warmup: Gradually increases batch size from 16M to 128M tokens for optimal training dynamics.
• Efficient Parallelism: Utilizes Ring Attention to minimize memory overhead for long sequences and padding optimization to reduce wasted computation.

Core Academic Benchmarks

(Tables showing benchmark results for General Tasks, Reasoning Tasks, and Mathematics & Coding Tasks are included here, mirroring the original input's tables.)

(Additional evaluation parameters link remains)

Getting Started with MiniMax-Text-01

(Code example for using MiniMax-Text-01 with Hugging Face transformers remains the same.)

Important Links

• Chatbot
• Online API
• Documentation

Conclusion

MiniMax-Text-01 demonstrates impressive capabilities, achieving state-of-the-art performance in long-context and general-purpose tasks. While areas for improvement exist, its open-source nature, cost-effectiveness, and innovative architecture make it a significant player in the AI field. It's particularly suitable for memory-intensive and complex reasoning applications, though further refinement for coding tasks may be beneficial.

The above is the detailed content of 4M Tokens? MiniMax-Text-01 Outperforms DeepSeek V3. For more information, please follow other related articles on the PHP Chinese website!