StoneforgeLABS

The research

Work with
something to say.

Published inquiry, practical experiments, and open-source work. The findings matter. So does how we arrived at them.

01 / Technical report

2026

AI systems · Gemma Rig

Running Gemma 4
locally.

Measured serving, evaluation, and Gemini CLI integration on a single RTX 5090. A practical examination of what local AI can deliver, with the setup, measurements, and limitations available for inspection.

Read the report Source code
Download report (PDF)
189 tok/s

Output speed

26B-A4B NVFP4
64k

Context window

Single RTX 5090
82.86%

MMLU-Pro

26B-A4B FP8
97→180

Output tokens/s

After expert offload

Reported batch-one decode under WSL2 with vLLM 0.25.1. Results describe these specific configurations; reproduction details and limitations are in the report.

02 / Open work

The work
behind the findings.

01

Stoneforge Labs repositories

Source code, reproducible tooling, and supporting project material.

Browse GitHub 02

Models & experiments

Explore the Stoneforge Labs profile on Hugging Face.

Hugging Face 03

Research profile

Publications and research interests under the Jayson Stone profile.

Google Scholar