RYIUK
RESEARCH LAB / 2026

Language models
on ordinary machines.

We test what runs on the hardware people already own. No supercomputer. What works there works anywhere.

HardwareOne ordinary PC
ModelsSmall, quantized
MeasureFive runs each
ResultPublished, failures too
01 / THE MACHINE

One test machine.

CPUAMD Ryzen 5 5600G, 6 cores, 12 threads
RAM30 GB
iGPUVega, on the same chip
dGPUGTX 1650, 4 GB, no tensor cores

We checked the GPU instead of assuming. The tensor core flag is false. That is the point: the hard case is the useful case.

02 / THE NUMBERS

What held, and what did not.

42%of tests held
66ablations run
9 / 4models on disk / ready
CPU27 tok/s
iGPU28 tok/s
GTX 165082 tok/s

The same small model, three devices. Five runs each. The middle bar is the GPU built into the processor.

Four of the nine model files are ready to serve. The rest are source copies, vision projectors or links.

One fix took a task from 0/8 to 8/8. We tried it on three more tasks and it failed on all three. We kept the failure.

03 / THE WORK

Four lanes.

InferenceSmall models on CPU, RAM, integrated and discrete GPUs.llama.cpp, 5 runs per cell, CPU only at 6 threads
ArchitectureInstrumented builds. Attention, quantization, routing.A fork with per-layer taps. Tap capture lives in one Windows-only call site
RetrievalRetrieval first. Weights only after the tests.RAG now, small fine-tunes later, then quantize
AgentsTools, rules and a shared log. Built for one job at a time.Typed tools, one append-only log, a gate that blocks headless destructive calls
How we check
Verifier1+1=2 proved in four steps. 1+1=3 rejected with the failed step shown.
Tool gate7 invariants. A destructive call from a headless agent is blocked, not asked.
Our own toolsWe filed two defects against them. A byte count and a rounding drift.
04 / OPEN CHANNEL

Work with us.

Local inference, retrieval and agent pipelines for teams running ordinary hardware.

support@daavile.com