Skip to content
SparQ

Frontier models

Every new model is an upgrade, not a rewrite.

New frontier models arrive every few months. With SparQ, taking one is a Release of your Project: evaluated on your own tests, approved by a second person, live in seconds, and reversible in one step. Your application keeps calling the same contract.

r40r41r42r43reasoningevals

Frontier models

Ask for a class of model. Let policy pick the deployment.

Agents never name a model. They ask for a class, and the Frontier Gateway chooses a deployment your workspace allows and each client's compliance profile permits, or refuses.

Model classes
classwhat agents use it for
reasoningMulti-step planning, analysis, careful answers
standardEveryday conversation and drafting
fastClassification, routing, short replies at low cost
codeWriting and reviewing code
visionScreenshots, diagrams and scanned documents
embedTurning knowledge into vectors for search
rerankOrdering search results by relevance

Providers today

Early access

Azure OpenAI and any OpenAI-compatible server, including self-hosted open-weight models, through the official provider SDKs.

More providers

Roadmap

Further frontier providers and managed model platforms, added to the registry without changing a single agent.

  • Retries with jitter and a circuit breaker per deployment
  • Fallback that stays inside the allowed list
  • Every routing decision recorded, with the region chosen
  • Deployments without stated training and retention terms are refused

Stay current

The model changed on Tuesday. Your app didn't.

Here is what moving a Project to a new model looks like in an ordinary week. Nobody opens your application's repository.

An example week of a model upgradeMonday release r41 is live. Tuesday a draft r42 moves the reasoning class to a new deployment. Wednesday evaluations pass. Thursday a second admin approves. Friday r42 is live, while runs already in flight finish on r41. The application does not deploy all week, and rollback is one step.MonTueWedThuFriyour applicationbuild 3.2.0no deploy, no code changeSparQ Project · prodreleasesr41 livereasoning: deployment Adraft r42reasoning → deployment Bevaluations passgolden · leak · injectionapprovedby a second adminr42 liveruns in flight finish on r41rollback: one step
Fig. 1  An example week. The release numbers and results are illustrative.

What the approver sees

A Release is an immutable bundle: models, instructions, agents, tools, guardrails, knowledge and the evaluation baseline. The approver sees exactly what changed, how quality and cost moved, and who wrote it. They cannot be the same person.

  1. Draft
  2. Validate
  3. Evaluate
  4. Approve
  5. Publish
  6. Roll back
Release diff, as the second admin reviews it (illustrative)
  release r41 → r42                       project: support-desk
  ───────────────────────────────────────────────────────────
  model_classes.reasoning
-   deployment: deployment-a  (region: eu-west)
+   deployment: deployment-b  (region: eu-west)
  evaluations
    golden      312 / 312 pass
    leak         48 / 48  pass
    injection    96 / 96  pass
  estimated cost per run     -18%
  author: priya   approver: marco   (author cannot approve)

Give your next AI feature a contract that outlives the model.

Tell us what you are building. We will show you SparQ on your use case and your residency rules.

Book a demoExplore the API

Early access: Built and tested. Available to early-access workspaces.