NHNH Applied
Home / Projects / QES

QES / Host-side model execution

Running model workloads beyond normal memory residency.

QES investigates whether very large language-model workloads can execute on a single 128 GB AMD system by managing model state and routed expert data rather than requiring full residency.

128 GB
single-system unified memory
~375 GB
Kimi K2.5 quantized model package tested
1T-class
model family scale demonstrated functionally
Windows
consumer workstation environment

The research question

Large mixture-of-experts models are normally discussed in terms of aggregate accelerator memory or multi-node deployments. QES asks a narrower engineering question: can the workload execute correctly on one 128 GB machine if the host controls what becomes resident, when it is needed and how routed expert data is streamed?

Claim boundary: QES demonstrates functional single-node execution and measured optimisation. It does not claim throughput parity with distributed accelerator clusters, universal model compatibility or production-chat performance.

What was demonstrated

  • Qwen3-235B-A22B quantized execution using host-managed expert streaming on a single machine.
  • Kimi K2.5 UD_Q2_K_XL, approximately 375 GB, executed on the same 128 GB class system.
  • Evidence-driven optimisation continued after functional success, including storage, CPU/GPU and memory-path investigation.
  • Runs are treated as engineering evidence: environment, faults, changes and measured results are preserved rather than replaced by a headline benchmark.

Why it matters

The result is most interesting where data locality, on-premise control or agentic workloads matter more than interactive chat speed. It explores a different trade space: using software-managed residency and commodity storage to extend what one local machine can attempt.