Design-partner pilot

Bring one expensive workload

We are recruiting a small group of design partners for a pilot starting in 2027. It begins with a benchmark request: you name one model, one GPU and one target, we plan the baseline run with you, then run the full loop and you keep everything it produces.

Pilot at a glanceDraft terms
Length
8 weeks
Scope
1 workload, 1 GPU target
Starts
2027, rolling
Partners
Small cohort
You keep
Engine, recipe, report
Draft pilot terms, subject to change.

What you get

Benefits for design partners

01

The full loop on your model

Profile, kernels, precision and serving, run on your workload.

02

Engineers alongside

Weekly reviews with the people building Turbokern.

03

Everything it produces

The engine, the recipe and the report are yours to keep and deploy.

04

A say in the roadmap

Your workload shapes what we build next.

05

Preferred terms

Design-partner terms when the platform launches.

06

A clear exit

If the pilot finds nothing worth keeping, the report says so.

The plan

Eight weeks, in four stages

  1. Weeks 1-2

    Baseline benchmark and profile

    Agree the objective, guards and eval set. Run the baseline benchmark. Profyle produces the opportunity map.

  2. Weeks 3-4

    Kernels and precision

    Kernl and Quantra runs, with weekly reviews of the candidate ledger.

  3. Weeks 5-6

    Serving

    Servex tunes against your traffic sample and packages the engine.

  4. Weeks 7-8

    Report and handover

    Final measurements, the recipe, and a review of what to deploy.

Fit

Who the pilot suits

The pilot works best for teams with a real inference cost and a clear target. We will tell you if it is not a fit.

  • You run a model in production or close to it.
  • You deploy on H100 or H200, or Jetson Orin for edge.
  • You have an evaluation set that defines acceptable quality.
  • Someone on your team can spend about two hours a week with us.
Not a fit yetTraining optimization: on the roadmap, not in the pilot.
Not a fit yetNon-NVIDIA hardware: Turbokern is built for NVIDIA GPUs only.
Not a fit yetModels you cannot run outside a vendor API.
Ask usMulti-tenant GPU sharing and safety-certified edge targets.

Benchmark request

Request a benchmark on your model

Name one model, the GPU it has to run on and the number it has to hit. An engineer reads every request and replies within five working days with a plan for the baseline run. Nothing runs until we agree the scope with you.

Use your company address.
Name and size, for example Llama 3.1 70B.
In ms, measured however you measure it today.
Tokens or requests per second per GPU.
A target such as p95 under 45 ms on one H100, or half the GPUs at the same throughput.
Or email pilot@turbokern.com

FAQ

Questions engineers ask

Is the pilot paid?

Terms are agreed per partner. Ask us when you request the benchmark.

Where do the runs happen?

On our GPUs, or in your cloud account if your data cannot leave it.

What if Turbokern finds nothing?

Then the report says so, with the measurements, and you have a profile of your model that you keep.

When does it start?

Cohorts start on a rolling basis in 2027.