Today, with the rapid development of technology, AMD once again leads the trend by launching Ryzen AI Max+395 based on the Zen5 architecture. This processor is not only a leader in the industry but also the first platform capable of running a 70 billion parameter AI large model locally. Now, AMD has announced a major upgrade, allowing the Ryzen AI Max+395 to support local operation of up to 128 billion parameter large models, further expanding its application potential in the AI field.
To achieve this performance, the Ryzen AI Max+395 still needs to be paired with 128GB of unified memory, of which 96GB must be allocated as VRAM. Additionally, the processor requires running in the Vulkan llama.cpp environment, providing developers with greater flexibility. The new upgrade makes the Ryzen AI Max+395 the first platform capable of running the 109 billion parameter Meta Llama4Sout model, which has a size of 66GB and supports advanced features such as Vision and MCP.
The core of this upgrade lies in the Mixture of Experts (MoE) mode, which activates only a part of the model at a time, significantly reducing resource consumption while maintaining performance. In actual testing, the Ryzen AI Max+395 can achieve a processing speed of 15 Tokens per second, which is quite impressive. It also supports the Mistral Large model, which has a size of 68GB and 123 billion parameters, as well as several smaller models, such as Qwen3A3B with 18GB and 300 billion parameters, and Google Gemma with 17GB and 270 billion parameters.
More surprisingly, the Ryzen AI Max+395 has made a significant leap in context handling, supporting a maximum context length of 256,000 Tokens, allowing it to process and analyze more complex data. While general non-large-scale models are sufficient with a context of 32,000 Tokens, this upgraded processor brings more possibilities. What used to require high-end equipment to achieve now can be easily realized on AMD's products.
In terms of price, the mini AI workstation equipped with the Ryzen AI Max+395 and 128GB of memory is now around 13,000 yuan, greatly improving its cost-effectiveness and significantly lowering the threshold for high-end AI applications.