About Bertholomus
BertholomusAI makes large open models run well on hardware you can own, and every result comes with evidence.
We publish reproducible recipes for distributed LLM inference and quantization on NVIDIA GB10 / DGX Spark systems. Each recipe includes pinned revisions, exact runtime settings, measured benchmarks, and known limitations, so you can rebuild the result yourself.
What we do
Multi-node deploymentsGLM-5.3 on 4× GB10, and DeepSeek V4 / V4.1 Flash on 2× and 4× GB10.
Quantization researchOriginal quants with clear authorship and full reproducibility.
Long context and speculative decodingWe test these features before we claim them.
Upstream contributionsWe send fixes back to the projects we build on.
Our standard
We credit every upstream model, runtime, kernel, and recipe we use.
We separate what is configured from what we have directly proven.
We publish the evidence and the limitations together with the headline numbers.
We never present third-party quants as our own work.
We use exact revisions and hashes, not tags that can change.
Who we are
BertholomusAI is a Bert Enterprises project, based in Los Angeles. We build and test everything on our own hardware.
Find us
Models and recipes are provided as-is. Check each upstream license before use.