AMD Ryzen AI Max+ 395, the only large model that locally runs 128 billion parameters, is a major upgrade

AMD Zen5 architecture Ryzen AI Max+ 395 is the first platform that can run a large 70 billion parameter AI model locally. Of course, it needs to be paired with 128GB of unified memory. Now,AMD has launched a major upgrade, allowing large models with 128 billion parameters to be run locally!

Of course, only Ryzen AI Max+ 395 can do this, and it is still necessaryPaired with 128GB unified memory, and allocate 96GB as video memory, and must be inVulkanllama.cpp environment.

The driver needs to be matched with version 25.10 RC 24, which is still in the testing stage, and will become the official version 25.8.1 early next month.

This makes Ryzen AI Max+ 395 the first platform that can locally run the Meta Llama 4 Sout large model with 109 billion parameters, with a volume of up to 66GB, a maximum of 17 billion active parameters, and support for Vision and MCP.

This is due to the model usingMixed Expert (MoE) Mode, only a part of the model is activated at a time to keep the performance available. The actual test of Ryzen AI Max+ 395 can achieve 15 Tokens per second.

At the same time, Ryzen AI Max+ 395 can also runMistral Large with 68GB volume and 123 billion parameterse, as for the 18GB, 30 billion parameter Qwen3 A3B, 17GB, 27 billion parameter Google Gemma, naturally not to mention.

Of course, Ryzen AI Max+ 395 also supports a variety of model sizes and quantization formats, including GGUF.

Equally important, the Ryzen AI Max+ 395 upgrade supports a larger context size, from the traditional 4096 Tokens to an astonishing 256000 Tokens, making it possible to analyze and summarize larger-scale documents and other data.

However, it is important to note here that Flash Attention and Q8 KV Cache must be turned on.

For general non-large-scale models, the context of 32,000 Tokens is enough, and Ryzen AI Max+ 395 provides new possibilities.In the past, such a large-scale context often required tens of thousands or hundreds of thousands of devices to implement.

The price of the mini AI workstation configured with Ryzen AI Max+ 395 and 128GB has been as low as about 13,000 yuan, which can be said to be quite cost-effective.