We put our tiny embedding model up against OpenAI and NVIDIA. It didn't blink.
crowsight is a small, offline embedding model that ships inside Crowkis. We didn't trust it on faith, so we made it compete with the biggest embedding APIs on the one job a semantic cache actually needs. Here's what happened.
There's a quiet insecurity that comes with shipping your own model. Everyone else is calling OpenAI's text-embedding-3 or a hosted NVIDIA endpoint, and here we are with a 22-megabyte file called crowsight baked into the binary. So we did the uncomfortable thing: we set up a head-to-head and let the giants swing first.
Three phrasings, one meaning, one cache hit. That's the whole game.
The test was deliberately real. We took genuine rephrasings, the same question worded differently, the way a real user or a real agent would, and measured how reliably each model recognizes them while safely ignoring genuinely different questions. In semantic caching that trade-off is everything: catch the paraphrase (you save money) without matching a different question (you'd serve a wrong answer).
On genuine rephrasings, crowsight matched the big hosted models, and beat the usual open-source picks.
The headline isn't 'we beat OpenAI', the top models all nail this task. The headline is that a small model you run yourself, with zero per-query cost and nothing leaving your box, sits in the same tier as embedding APIs that bill you forever. crowsight also had the cleanest separation of the whole field: rephrasings scored high, different questions scored near zero. Clean separation is what keeps a cache from confidently serving the wrong answer.
The bottom line
If your semantic cache phones an embedding API for every lookup, you're paying a tax to understand questions you've already answered, and shipping every prompt off-box to do it. crowsight makes that unnecessary. It's bundled, it's offline, it's fast, and on the job that matters it keeps pace with models a hundred times its size.