Ship any model to every device you support.

Octomil packages one model path across CoreML, LiteRT, MLX, ONNX, and browser runtimes, releases it by device cohort with holds and rollback, and falls back to cloud only when the device cannot deliver. Inference that stays on-device is never metered.

Free while traffic stays on-device. First response in under five minutes. Need a walkthrough instead? Request a demo.

Main.kt
import com.octomil.sdk.Octomil

val client = Octomil(publishableKey = "oct_p12....")
client.initialize()
val response = client.responses.create(
    model = "gemma3-1b",
    input = "What's the capital of Canada?"
)

println(response.outputText)
> The capital of Canada is Ottawa.

Take the shortest path to evaluation

Use the route that matches where your team is today.

Run locally

Validate the local path on your own hardware before you commit to rollout policy.

Open local setup

Compare the options

Where Octomil fits against cloud gateways like OpenRouter and LiteLLM, and where it does not.

Read the comparisons

See your savings

The same meter that runs in production: what your traffic costs through Octomil versus all-cloud.

Open the savings calculator

The release policy is legible

Runtime preference, device targeting, rollout state, and fallback rules live in one place, so what ships to which cohort is inspectable rather than buried in client code.

Decision inputs for `mobile-default@r19`

Example policy fields for a routine mobile cohort.

Device cohortiPhone 17 Pro
Preferred runtimeCoreML / Apple NPU
Policy versionmobile-default@r19
Fallback ruleCloud only on constraint
A request like support-ticket summarization can inherit this policy, stay on-device by default, and still fall back safely when hardware or quality limits show up.

One model path across every runtime you ship

Package once, release by cohort, and keep the production path legible.

Packaging

One packaging path across CoreML, LiteRT, MLX, ONNX, and browser runtimes.

Release control

Roll out by cohort, hold changes, and roll back without rebuilding the model path.

Fallback

Keep routine inference local and use cloud only when hardware or quality requires it.

Keep the review path straightforward

When procurement, privacy, or security requirements show up, the platform should still feel manageable.

Enterprise identity

Support for SSO, SCIM, scoped access, and audit history.

Deployment posture

VPC and privacy-sensitive deployment options when review requirements show up.

Operational review

Security materials, deployment planning, and rollout review support.

Ship one request this afternoon

Grab a key, point the SDK at a small model, and watch the router keep it on-device. When procurement or rollout planning shows up, we will walk through your actual workload.