Deployment

One command. A private endpoint you own.

Your tailored model is served behind a private, OpenAI-compatible endpoint at https://api.basisresearch.tech/v1. Point your existing client at it, swap the base URL and model name, and keep the code you already wrote.

01The endpoint

An API your engineers already know

Kethra's deployment is OpenAI-compatible. Your engineers do not learn a new SDK. They update two lines.

OpenAI-compatible chat and completions endpoints
Streaming responses supported out of the box
Private by default, accessible only with your API key
Served over HTTPS with TLS 1.3
JSON schema-validated request and response format
Versioned model names so integrations do not break on upgrade

client.ts

import OpenAI from "openai";

const kethra = new OpenAI({
  baseURL: "https://api.basisresearch.tech/v1",
  apiKey: process.env.KETHRA_API_KEY,
});

const res = await kethra.chat.completions.create({
  model: "your-house-v3",   // your tailored model
  messages: [
    { role: "user", content: "Draft a renewal note." },
  ],
  stream: true,
});

for await (const chunk of res) {
  process.stdout.write(
    chunk.choices[0]?.delta?.content ?? "",
  );
}

02Versioning

Per-version endpoints and instant rollback

Every time a new version of your model is deployed, a stable per-version endpoint is created alongside the primary one. Your integration can pin to a specific version whilst you validate, then cut over when you are ready.

Rollback is one command. If a newly deployed version underperforms in production, you revert immediately with no downtime and no rebuild.

What versioning gives you

Every deployment is versioned automatically, with a stable per-version endpoint.
Roll back to any previous version with a single command, with no downtime.
Canary traffic splits let you test a new version on a fraction of traffic before a full cut-over.
The deployment log records who deployed what and when, with a tamper-evident audit trail on Enterprise.

03Hosting

On our infrastructure or yours

Where the model runs is your decision. All three options use the same endpoint format and the same Kethra CLI.

Starter and Atelier

Kethra-managed

Your endpoint runs on isolated Kethra infrastructure in the United States. Versioning, rollback, scaling and uptime monitoring are included. You point your client at the URL and nothing else changes.

Atelier and Enterprise

Your own cloud or VPC

The model is deployed into your own cloud account or virtual private cloud using Kethra's deployment tooling. The endpoint is private to your network, and the weights never leave your environment.

Enterprise

Fully on-premises

For teams that cannot allow model data outside their own perimeter, training and serving run entirely on-site. Kethra engineers assist the initial setup and each refinement cycle. The weights stay inside your estate.

terminal

# Deploy to a private Kethra-managed endpoint
kethra deploy --private

# Deploy into your own VPC
kethra deploy --private --target vpc://your-account/us-east

# Deploy on-premises (Enterprise)
kethra deploy --private --target onprem://your-host

# Output in all cases:
#   Endpoint: https://api.basisresearch.tech/v1 (or your internal URL)
#   Model:    your-house-v3
#   Version:  3  (previous: 2, rollback: kethra rollback)

04Scaling and latency

Built to handle production load

Kethra-managed endpoints scale horizontally under load. Capacity is monitored continuously and scaled ahead of demand during regular business hours. For high-throughput workloads we offer dedicated capacity on request.

Latency

Latency depends on model size and prompt length. Distilled models, which are smaller and faster weights derived from the tailored model, typically offer lower latency with modest accuracy trade-offs. We report the latency profile of every deployment alongside the eval results so you can make the right choice for your workload.

SLAs

Uptime SLAs are available on the Atelier and Enterprise tiers. Current status and historical availability are published at basisresearch.tech/status.

05Distillation and export

Smaller weights. Yours to keep.

Distillation

Where the full tailored model is larger than your latency or cost requirements allow, Kethra can distil it to a smaller, faster set of weights. The distilled model is derived from your fine-tune, not from a generic smaller base, so it retains the voice and standard alignment of the original. Every distilled version is scored on the same eval before deployment.

Exporting weights

Your weights are yours to export at any time, in a standard format compatible with common open-weight runtimes. Export does not terminate your Kethra subscription. Use the console, or run the command alongside.

terminal

# Distil the tailored model to a smaller variant
kethra distil --from your-house-v3 --size 7b

# Output:
#   Distilled: your-house-v3-7b
#   Eval:      voice 87  accuracy 81  standard 93
#   vs full:   voice 89  accuracy 84  standard 95
#   Latency:   ~40% lower p50

# Export weights in safetensors format
kethra export --model your-house-v3 \
              --format safetensors \
              --out ./weights/

# Weights written to ./weights/your-house-v3/

Ship a private model in one command.

Your tailored model, behind an endpoint your engineers already know, on infrastructure you choose.

Start freeRead the API docs