Editorial illustration for AMD Ryzen AI Max+ runs 122B‑parameter models locally with 128 GB UMA
AMD Ryzen AI Max+ runs 122B‑parameter models locally...
The era of local AI inference has just crossed a new threshold. AMD’s Ryzen AI Max+ processor, paired with a staggering 128 GB of unified memory, now runs a 122-billion-parameter model entirely on your desk. No cloud.
No compromise. The headline is simple: 122B-class weights are no longer the exclusive domain of server racks. They fit in a PC.
The numbers tell a story of surprising efficiency. A 9B model? Trivial, 100% GPU offloaded, blazing fast.
The 35B model, despite its size, actually outpaces the 9B. That’s the magic of Mixture-of-Experts architecture: only 3 billion of its 35 billion parameters activate per token. Speed comes from selectivity.
Then there’s the 122B behemoth, 76 GB of raw weight. It exceeds the 64 GB GPU-accessible memory ceiling, so Ollama automatically splits the load: 61% on the GPU, 39% on the CPU. It’s a hybrid dance, not a bottleneck.
The implications are immediate. You can run a 122B-parameter model locally, on a single chip, with no cloud dependency. Token generation speeds vary by architecture, MoE models punch above their weight class.
And with BIOS-level GPU memory allocation tweaks, even more layers can shift to the GPU. The hardware is ready. The software manages the rest.
AMD Ryzen™ AI Max+ processors with AMD Radeon™ 8060S integrated graphics change that equation. With up to 128 GB of unified memory shared between CPU and GPU, the entire memory pool is accessible for AI workloads — enabling models with 100 billion or more parameters to run on a single system, with no second card and no cloud bill.
The 122B-parameter model, once confined to data-center clusters, now runs on a single desktop CPU, 76 GB of weights straddling GPU and system memory, orchestrated by Ollama’s automatic loader. That’s not a compromise; it’s a threshold crossed. The 35B MoE model already outruns its denser 9B sibling by activating only 3 billion parameters per token.
Now imagine what happens when BIOS allocations shift more layers to the GPU. The Ryzen AI Max+ doesn’t just inch toward local large-model inference. It bulldozes the gate.
And with 128 GB of unified memory, the ceiling isn’t 122B, it’s whatever the next weight set demands.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Semantic Search Model Assigns Class Labels and Confidence Scores to Critiques
From Archives Hotz warns AI coding agents could be costly despite 10x productivity boost
Common Questions Answered
How does the AMD Ryzen AI Max+ handle 122-billion-parameter models with 128 GB of unified memory?
The Ryzen AI Max+ runs 122B-parameter models entirely locally by distributing 76 GB of weights across both GPU and system memory, orchestrated by Ollama's automatic loader. This unified memory architecture eliminates the need for cloud-based inference while maintaining performance, representing a significant shift in making large language models accessible on desktop computers.
What is the performance difference between the 9B and 35B MoE models on the Ryzen AI Max+?
The 35B MoE (Mixture of Experts) model actually outperforms the denser 9B model despite being larger, because it only activates 3 billion parameters per token rather than processing all weights. This efficiency demonstrates how model architecture, not just parameter count, determines real-world performance on the Ryzen AI Max+ platform.
Why is running 122B-parameter models locally on the Ryzen AI Max+ significant for AI inference?
Previously, 122-billion-parameter models were confined to data-center clusters and required cloud computing resources. The Ryzen AI Max+ crosses a critical threshold by enabling these large models to run on a single desktop PC without compromise, democratizing access to advanced AI inference and eliminating latency and privacy concerns associated with cloud-based processing.
How does GPU offloading work for different model sizes on the Ryzen AI Max+?
Smaller models like the 9B parameter version achieve 100% GPU offloading and run at blazing-fast speeds, while larger models like the 122B distribute their weights between GPU and system memory. The article suggests that further BIOS allocation adjustments could shift more layers to the GPU, potentially improving performance for even larger models.
Further Reading
- AMD Ryzen™ AI Max+ Upgraded: Run up to 128 Billion parameter LLMs on Windows with LM Studio — AMD
- AMD Ryzen™ AI Max+ AI PCs Deliver Exceptional Intelligence Right — AMD
- AMD brings 128B LLMs to Windows PCs with Ryzen AI Max+ 395 — SecurityBrief New Zealand
- How To Run OpenAI's GPT-OSS 20B and 120B Models on AMD Ryzen AI Radeon — AMD
- Review: Minisforum MS-S1 Max | AMD AI Max+ 395 with 96GB of VRAM — AkitaOnRails
