Running a 1-Trillion-Parameter LLM on One 128 GB Ryzen AI MAX+ PC

QES demonstrates execution of the ~375 GB Kimi K2.5 Mixture-of-Experts model on a single AMD Ryzen AI MAX+ 395 Windows PC with 128 GB unified memory.

Read the QES v2.7 White Paper GitHub

Why this matters

AMD publicly demonstrated the 375 GB Kimi K2.5 workload using a four-node cluster of Ryzen AI MAX+ 395 systems, each with 128 GB of memory.

QES investigates a different deployment question:

Can that model class remain persistently available for local execution on one 128 GB system instead of requiring four coordinated machines?

QES demonstrates that it can — at substantially lower throughput than the four-node cluster, but with the workload retained on one Windows machine, leaving three equivalent systems available for other work.

Published Kimi result

~375 GB Kimi K2.5 model
128 GB single-node unified memory
0.432091702 t/s sustained over 128 generated tokens
+80.87% versus frozen CPU baseline
61,440 GPU expert executions
7,680 authoritative correctness checks
129 / 129 output token IDs matched reference
0 GPU errors

Qwen → Kimi cross-family portability

QES was first demonstrated with Qwen3-235B and was then adapted and qualified for the substantially larger Kimi K2.5 architecture.

Model Routed MoE layers Experts / layer Routed-expert payload
Qwen3-235B 94 128 137.29 GB
Kimi K2.5 60 384 367.30 GB

Kimi represented approximately a 2.67× increase in routed-expert payload while also changing model family and expert topology.

The published claim is deliberately limited: QES has demonstrated cross-family portability across two materially different MoE architectures. It does not claim universal model agnosticism.

Model integrity and high-assurance use

QES treats model integrity and evidence preservation as part of the execution architecture rather than as post-run checks.

The public work includes pre-execution verification of qualified expert stores, protection against unintended modification, continuous downstream correctness checking, evidence-preserving qualification, and fail-closed behaviour when required conditions are not satisfied.

These properties may be particularly relevant to medical, legal, defence, industrial and other controlled on-premises or high-assurance environments, where knowing which model data was executed and whether the qualified execution path remained valid may matter alongside raw inference speed.

QES does not claim regulatory, medical, legal, defence or safety certification.

Test platform

The published Kimi validation used an AMD Ryzen AI MAX+ 395 with Radeon 8060S, 128 GB unified memory, Windows 11 Pro, and ordinary internal NVMe storage.

The accelerated storage path used a Samsung SSD 990 EVO Plus 2 TB authoritative expert store and a YMTC PC41Q-2TB-B verified secondary expert-store replica.

Public disclosure boundary

The public QES materials intentionally do not disclose reconstruction- enabling implementation details including exact scheduling ratios, expert-record formats, internal state machinery, cache/eviction implementation, memory-transfer implementation, launch environment, source paths or build recipe.

Technical White Paper — DOI 10.5281/zenodo.22730031 NH Applied QES on GitHub AMD Four-Node Reference