An overnight tune-up for AI models
Refresh consolidates how your model reads its memory — overnight, on your infrastructure. It wakes reading 25–45% less, with a certificate proving every ability you care about survived the night.
Your model never leaves your boundary. Nothing inside it is rewritten. Every change is reversible.
How it works
Refresh never retrains and never edits your model’s weights. It tunes a single thing: which parts of its memory each attention pathway is allowed to read.
Evening
Your fine-tuned, private model, on your own machines or cloud. Weights stay where they are — the pipeline visits the model, not the reverse.
Night
Refresh watches how the model actually reads, keeps the pathways that carry real signal at full reach, and gently narrows the rest — over many cycles, the way sleep consolidates a brain.
Morning
A new serving configuration that reads far less memory — delivered with a before-and-after ability certificate for this exact deployment.
The science
Foveated is named for the trick your own vision uses — and Refresh applies it with the care of sleep and the patience of annealing.
Your eye sees sharply only at its center — the fovea — and VR headsets exploit exactly that to render 10× cheaper with no visible loss. Refresh gives a model’s attention a fovea: full reach where it matters, light reach everywhere else.
Every night your brain downscales its connections, keeps the strong, and lets the weak go — that is how capacity returns by morning. Refresh runs the same rule over a model’s attention pathways, cycle by cycle.
Cut a model’s memory in one stroke and abilities crack. Refresh lowers the budget a little each cycle — the blacksmith’s slow cool — so the model settles into its strongest configuration.
The certificate
Here is the uncomfortable fact our research demonstrated: a model can lose its long-range memory entirely and still pass the industry’s standard benchmarks.
So every Refresh ends with a hidden exam the tuning never sees — recall a fact planted a thousand words back, repeat an exact number, keep track of who’s who — with questions minted fresh for each certification. Pass, or it doesn’t ship.
You specify the abilities your workload lives on. We mint the questions. That split is what makes the certificate worth more than a self-graded benchmark.
Ability certificate
Issued per deployment, bound to its exact configuration. Exam held out from everything that optimizes.
The numbers
of memory reads removed at full measured ability, across three research-scale models
abilities lost on the exam at the certified operating points — the biggest model took the deepest cut
taken off one real-shaped 16-GPU fleet’s memory bill in our worked example
the saving grows with context: 65% of full reads at 2k tokens trends toward 45% at one million
Pricing
The certificate measures the saving; the invoice is a fraction of it. If the night doesn’t pay for itself in its first month, it’s free.
The tune-up
From $36k /fleet/year
anchored at 25% of verified savings
The fleet subscription
Custom
models drift — sleep is a schedule, not an event
The certificate alone
Per deployment
for any efficiency claim — not just ours
Questions
No. Weights are never touched. Refresh tunes a reading policy — which parts of memory each attention pathway may read. A wrong setting is undone by widening the policy; nothing is ever lost.
No, and we won’t claim it. Refresh makes models cheaper, provably not worse: the certificate shows abilities preserved, and the savings show up on the serving bill. Speed, concurrency, and affordable context length improve; the model’s knowledge doesn’t change.
Yes — that’s the design center. The pipeline visits your model inside your boundary; nothing about your weights, prompts, or data leaves it. The certificate is the only thing meant to travel.
Certified at research scale: three models from 410M to 6.9B parameters, with preregistered exams and public methodology. Production-scale certification is the current program — ask us where it stands the day you read this.
You specify the abilities in your workload; exam instances are minted fresh at certification time so nothing can be tuned to them. Broad reading-quality floors ride along on every certificate to catch collateral damage outside your spec.
Design partners — now enrolling
We’re enrolling a small number of design partners with private, long-context fleets — regulated and on-premises teams first. One model, one night, one certificate: that’s the whole first meeting.
hello@getrefreshed.ai