AMD Breaks New Ground with Ryzen AI Max+ 395 Processor

AMD has once again pushed the boundaries of processing power with its latest Ryzen AI Max+ 395 processor. Based on the Zen5 architecture, this cutting-edge chip now supports local operation of artificial intelligence models with up to 128 billion parameters—a significant leap from its previous capability of handling 70 billion parameters.

Technical Specifications and Requirements

To achieve this remarkable performance, the processor requires:

• 128GB of unified memory, with 96GB allocated as VRAM
• Operation in the Vulkan llama.cpp environment, providing developers with enhanced flexibility

The upgrade positions the Ryzen AI Max+ 395 as the first platform capable of running Meta's Llama4Sout model (109 billion parameters, 66GB size) which supports advanced features including Vision and MCP capabilities.

Innovative MoE Technology

The breakthrough comes from AMD's implementation of Mixture of Experts (MoE) technology. This approach activates only a portion of the model at any given time, dramatically reducing resource consumption while maintaining high performance levels. In benchmark tests, the processor achieves:

• Processing speeds of 15 tokens per second
• Support for multiple large models including:
  • Mistral Large (68GB, 123B parameters)
  • Qwen3A3B (18GB, 300B parameters)
  • Google Gemma (17GB, 270B parameters)

Enhanced Context Handling

The Ryzen AI Max+ 395 shows particularly impressive gains in context handling:

• Supports context lengths up to 256,000 tokens
• Standard non-large-scale models typically require only 32,000 tokens This expanded capacity enables processing and analysis of significantly more complex datasets than previously possible on consumer hardware.

Accessibility and Pricing

Perhaps most notably, AMD has managed to bring this high-end performance to a more accessible price point:

• Complete mini AI workstation solutions featuring the Ryzen AI Max+ 395 and 128GB memory are available for approximately 13,000 yuan
• This represents a substantial reduction in the cost barrier for advanced AI applications

Key Points:

1. AMD's Ryzen AI Max+ 395 now supports local operation of 128B-parameter AI models
2. Requires 128GB unified memory (96GB VRAM) and Vulkan llama.cpp environment
3. Uses innovative Mixture of Experts (MoE) technology for efficient processing
4. Handles context lengths up to 256K tokens
5. Priced at ~13K yuan for complete workstation solutions