Skip to content
Let's talk

Futurum Research

Insight · Market Coverage

AI Platforms5 min read

AMD Expands Windows AI Limits With 128B Parameter Model Capability

Research Director & Practice Lead, Intelligent Devices

Updated

The short answer

Olivier Blanchard, Research Director at Futurum, shares his insights on AMD’s Variable Graphics Memory upgrade for the Ryzen AI Max+ 395, enabling up to 128B parameter models with full vision and MCP support on Windows devices.

Futurum's Olivier Blanchard,

Cite this

Olivier Blanchard, The Futurum Group, "AMD Expands Windows AI Limits With 128B Parameter Model Capability," August 13, 2025. https://trial.futurumgroup.com/insights/amd-expands-ryzen-ai-max-with-128b-parameter-model-capability/

AMD Expands Windows AI Limits With 128B Parameter Model Capability

Analyst(s): Olivier Blanchard
Publication Date: August 13, 2025

AMD’s Variable Graphics Memory upgrade for the Ryzen AI Max+ 395 enables up to 128 billion parameter LLMs on Windows with vision and MCP support, bringing data center-level AI to thin, light devices.

What is Covered in this Article:

  • AMD’s Variable Graphics Memory upgrade for Ryzen AI Max+ 395 enables up to 128 billion parameter LLMs on Windows.
  • Integration with Vulkan llama.cpp and LM Studio supports advanced vision and MCP capabilities.
  • Expanded context length up to 256,000 tokens supports complex, token-heavy workflows.
  • The upgrade positions AMD as the only vendor offering full-stack, cloud-to-client AI workload capability in thin and light systems.

The News: AMD has rolled out a significant update to its Variable Graphics Memory (VGM) tech, letting the Ryzen AI Max+ 395 (128GB) processor handle large language models with up to 128 billion parameters in Vulkan llama.cpp on Windows. This update will be part of the upcoming Adrenalin Edition 25.8.1 WHQL drivers, giving thin and light Windows systems the ability to use up to 96GB of VGM for AI tasks.

With this upgrade, the Ryzen AI Max+ 395 becomes the first Windows AI PC processor to run Meta’s Llama 4 Scout 109 billion (17 billion active) with full vision and Model Context Protocol (MCP) support. It solidifies AMD’s position as a go-to platform for running models from 1 billion to 128 billion parameters locally through llama.cpp.

AMD Expands Windows AI Limits With 128B Parameter Model Capability

Analyst Take: This jump in AMD’s Variable Graphics Memory marks a significant inflection point for running AI locally on Windows, as it breaks past limits that once limited workloads of that size to data center environments. By enabling 96GB of dedicated graphics memory on the Ryzen AI Max+ 395, AMD makes it possible to run anything from small assistants to massive vision models, all while supporting the long context lengths needed by advanced AI agentic solutions.

Expanded Model Capacity

Now able to handle up to 128 billion parameter models, the Ryzen AI Max+ 395 can run tasks requiring significant VRAM blocks without falling back on slower shared memory. This includes heavyweight models like Mistral Large and Llama 4 Scout (which must load the full 109 billion parameters even if only 17 billion are active once).

These models’ ability to achieve speeds of up to 15 tokens per second is proof that Ryzen AI Max+ 395 can handle large on-device AI workloads smoothly. It further closes the gap between consumer-class and enterprise-class AI hardware by bringing massive model sizes to devices with more portable form factors.

Context Length Advantage

The chip supports context lengths up to 256,000 tokens with Flash Attention ON and KV Cache Q8, making it ideal for token-heavy work. AMD’s demos – which include summarizing its quarterly reports through MCP or processing entire ARXIV research papers – demonstrate that it can handle over 19,000 and 21,000 tokens (respectively) in a single run. The Ryzen AI Max+ 395 can enable a PC to keep deep conversation history and context going across long sessions, as well as enable advanced agent workflows, and do so securely (as in: inside of an enterprise firewall) and without the need to consume cloud-based resources.

Flexible Quantization and Precision

The upgrade works with a wide range of quantization formats in llama.cpp and GGUF, from lightweight Q4 K M setups to high-precision 16-bit models. While Q4 K M is great for general use, it can also run Q6 or Q8 for tasks that need extra accuracy, like coding or vision-based inference. The architecture even supports running models like Google Gemma 3 27B in FP16, using the high memory bandwidth of the Strix Halo platform. This flexibility ensures users can balance speed, accuracy, and model size to match specific workload demands without hardware limitations.

MCP and Agentic Workflow Enablement

With full MCP and tool calling support when paired with compatible software, the Ryzen AI Max+ 395 presents as a critical and timely missing link in the effort to enable on-device AI agents. It can also handle the extra token load from MCP documentation and tool call returns, which can easily add tens of thousands of tokens. This keeps multi-step, multi-tool workflows stable and responsive – an important distinction against smaller systems that would struggle with that kind of workload. As MCP gains traction with developers like Meta, Google, and Mistral, this processor is set to become a strong local hub for complex, context-rich AI operations. Lastly, as the AI ecosystem begins to transition from primarily a training focus to the scaling of inference, processors like the Ryzen AI Max+ 395 provide PC users with a solid future-proofing foundation that ensures maximum sustained ROI from their PC investment, even this early in the refresh cycle.

What to Watch:

  • Adoption of AMD’s Variable Graphics Memory upgrade by developers targeting large-scale local LLM and VLM deployments.
  • User uptake of extended context length capabilities for MCP-driven, token-heavy workflows.
  • Performance scaling across various quantization levels and precision modes in real-world scenarios.
  • Competitive responses from other vendors offering local AI model execution on Windows devices.

See the complete announcement on AMD’s Variable Graphics Memory upgrade for Ryzen AI Max+ 395 on the AMD website.

Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.

Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.

Other insights from Futurum:

AMD Q2 FY 2025 Sales Beat Offset by MI308-Linked EPS Decline

AI PCs, Ryzen, and the Next Frontier in Personal Computing – Six Five On The Road

AMD Expands Telecom Role as Nokia Selects EPYC for 5G Cloud Platform

Image Credit: AMD

Published by Futurum.

About the author

Olivier Blanchard

Olivier Blanchard

Research Director & Practice Lead, Intelligent Devices

Olivier Blanchard has extensive experience managing product innovation, technology adoption, digital integration, and change management for industry leaders in the B2B, B2C, B2G sectors, and the IT channel. His passion is helping decision-makers and their organizations understand the many risks and opportunities of technology-driven disruption and help them leverage innovation to build stronger, better, more competitive companies. A trusted source of analysis and insights on digital business and digital innovation, Olivier is also the best-selling author of Social Media ROI: Managing and Measuring Social Media Efforts in Your Organization, and co-author of Building Dragons: Digital Transformation in the Experience Economy and Human/Machine: The Future of our Partnership with Machines. Blanchard is based in Greenville, South Carolina.

Coverage on AMD
  • Analyst Insights274
  • Research Reports1
  • Documents1
All AMD coverage
See the full picture in the platform

Vendor scoring, the benchmark data behind these findings, and the analyst team that produced them.

Open the platform (opens in a new tab)Become a client
The Weekly

Subscribe to The Weekly

The week's market signals, one email, no filler. Unsubscribe any time.