Home / Lab / Research

We measure before we ship.

Benchmarks and experiments run on our own cluster. Best cases are labelled as best cases.

Inference

Replicas vs tensor-parallel

2 small > 1 big

On the same GPUs, two smaller vLLM replicas served more requests than one large tensor-parallel split. We size client hardware with this in mind.

Best case measured: +136% throughput. Typical gains are lower and depend on the model and load; we quote the low end to clients.

Agents & the web

CTX: the web for models

~85% fewer tokens

Real web pages converted to CTX cost agents far fewer tokens than raw HTML. 100-site test: 81 pass, 13 partial, 4 fail, 2 timeouts. We publish the failures too.

Search

Sorted: our own index

3.7M+ pages

A curated search index built and served from our lab, so our agents get web search without a paid search API.

Model R&D

Pruned + quantized coder model

REAP + AWQ

Expert pruning plus 4-bit AWQ on a large coding model, so it fits on hardware a business can actually buy. Ongoing; numbers go here once they are reproducible.

Model R&D

Tokenizer adaptation

tokopt

Fitting a model's vocabulary to a domain so the same text costs fewer tokens. Ongoing research; used in model-training engagements.

Start with a free exploration call.

Tell us what you're trying to do and what's in the way. If AI fits, we'll show you where. If it doesn't, we'll say that too.