PRIVATE MODELS // YOUR CLOUD // YOUR WEIGHTS
Private frontier deployments still rent you someone else’s model, one they can retire, reprice, or change under you. We build you a specialized model that’s at par with the frontier on your task, runs in your cloud, and is yours to keep.
EXAMPLE DEPLOYMENT / PRIVATE INFERENCE
Legal model
Distilled.Trained.Hosted.Owned.
On (your) premises.
Training, evaluation, and inference stay in your environment. You own the model and control each release.
[01] HOW IT WORKS
Documents, contracts, processes, secrets.
The teacher model runs in your cloud too.
Google Cloud
AWS
AzureMeasured on your real work.
Private endpoints for your apps.
You now have your own
private AI model.
[02] CONTINUOUS IMPROVEMENT
Our work doesn’t end at delivery. We stay on to keep improving your model as your needs change: new data, new tasks, new teams.
As your teams use it, prompts and results can be saved and scored, automatically or by your reviewers. Approved results become training data for the next version. It all runs in your cloud, and every release passes evals and your approval.
Prompts and results are saved in your cloud.
Scored automatically or by your reviewers.
Approved results become post-training data.
New version ships only after it passes evals and you approve.
# release.log (example) v1.0 baseline ✓ evals passed ✓ approved v1.1 +12,400 examples ✓ evals passed ✓ approved v1.2 +9,800 examples ✓ evals passed ✓ approved v1.3 collecting
[03] THE POC
// The clock starts at kickoff: contract signed, onboarding done, cloud access granted.
[04] FOUNDERS
We’ve been building specialized models for years: cybersecurity and trading models for our own internal use, and a video editing model with Eddie AI. In every case the data was too sensitive to send anywhere else.
For Eddie, we distilled a 2.8-trillion-parameter frontier model into a 9-billion-parameter model for a single objective. That’s over 300× smaller.
To our surprise, it reached parity with the frontier on that objective. And it runs on a single GPU, even a laptop’s. No GPU cluster required.
Then companies started asking: can you build a specialized model for us? So we turned it into a service, available to any enterprise, at a fraction of the cost of a comparable frontier model, with a fast turnaround.
Discuss a POCHEAD-TO-HEAD VS. 2.8T FRONTIER MODEL
Share of blind head-to-head comparisons where the 9B model was preferred, on held-out real projects.
50% is as good as it gets: the two models are indistinguishable. 47% is effectively parity.
“Customers are not opposed to AI training; they are opposed to losing control of the knowledge they create.”Shamir Allibhai, in CineD ↗
CO-FOUNDER
George co-founded sFOX, a crypto prime brokerage that has processed over $600B in transaction volume through algorithmic trading. He built our in-house cybersecurity and trading models.
PREVIOUSLY BACKED BYKhosla Ventures · DCG · Tribe Capital · Social Capital · Y Combinator · Boost VC
CO-FOUNDER
Shamir built and sold video AI companies, most recently to Meta, where he oversaw video infrastructure as its first PM. At Eddie AI, he led the distilled editing model.
PREVIOUSLY BACKED BYOffline Ventures · Ashton Kutcher’s Sound Ventures · Gradient, Google’s AI fund · The Production Board
[05] FAQ
[06] START WITH ONE WORKFLOW
Tell us about one workflow. We’ll discuss your needs, scope a proof of concept, and define how it will be evaluated in your cloud.