RESPONSE TIMEEvery request.
Every request.
Every millisecond.
See how response times vary across a session, from the typical request to the slowest few.
Sample window60min
Mean p50157ms
Mean p99335ms
SMALL MODELS FOR AI-FIRST PRODUCTS
Your product.
In plain language.
Turn user requests into product actions, without per-token fees. Small models trained for your app run directly on your users’ devices.
Explore the samples02 PERFORMANCE
Performance,
in context.
Response time, task quality, and model size.
Explore a few of the things worth measuring.
SAMPLE DATA
Interactive visual samples. Minifield performance results are still to come.
QUALITY & SIZEFind the
Find the
useful tradeoff.
Compare task success with the space a model needs. Explore where a smaller footprint meets the quality your product needs.
Selected sampleModel C
Task success94.1%
Relative size38%
Task success (%) / relative model size (%)
TRAINING PROGRESSFollow the
Follow the
learning curve.
Track a training run alongside held-out examples. Watch for improvement that carries beyond the training data.
Sample steps2,000
Training loss0.25
Validation loss0.43
03 FOR PRODUCT TEAMS
Reduce onboarding
and documentation.
Let users complete common tasks in their own words while they’re still learning your product.
Get users productive sooner.
Reduce the need for walkthroughs. Give new users a direct way to get work done before they’ve learned where every control lives.
Put the whole product to work.
Help users find and use the features you’ve already built. Turn familiar requests into supported actions, right where they’re working.
Reduce how-to support.
Give your team fewer routine workflows to explain. Let users filter, organize, and change views in their own words.
04Built for product economics
Grow your user base.
Keep inference off the bill.
Make AI part of every session without a per-user token bill. Small models run on your customers’ devices, keeping hosted inference charges out of local interactions.
No per-token fees refers to local inference. Model training, integration, distribution, and any cloud fallback still have costs.
05 PERMISSIONS & PRIVACY
Keep your permissions.
Protect customer data.
Bring AI into your product with the access controls you already use. Keep prompts and context on the user’s device during inference.
YOUR CUSTOMER’S DEVICE
Their request“Show work in progress.”
Small modelPURPOSE-BUILT
Your appReady when you are.
PRODUCT
Built around your app.
Use a small model trained for your product’s vocabulary and supported actions. Focus its training on the tasks your customers need.
PRIVACY
Protect customer privacy.
Process requests on the user’s device, without sending prompts or context to an external AI provider.
PERMISSIONS
Use existing permissions.
The AI can only access what the user can access and perform actions they’re already allowed to take.
06Our point of viewBuild small.
Build small.
Believe deeply.
Inspired by the idea at the heart of Asimov’s Foundation: knowledge, distributed widely, gives small groups the power to shape what comes next.
Independent by design.- 01
Power belongs with people.
The device in your hands should work for you. We believe intelligence should live there, too.
- 02
Small is a deliberate choice.
A focused model should know your product’s job deeply. Keep its footprint small enough to make broad access practical.
- 03
Intent is the interface.
Users bring the goal. Your product should help them get there, with clear actions and an easy way back.
- 04
Privacy starts with proximity.
Keeping inference on your device means your instructions can stay there. That’s a design decision we make at the beginning.
- 05
Capability should be yours to keep.
We believe in tools that keep working. Local intelligence makes room for ownership, offline use, and life beyond the meter.
SMALL MODELS FOR AI-FIRST PRODUCTS.