AI Inference Benchmark: Keln vs Fireworks, Baseten, Modal, DeepInfra & Z.ai
ProductLite mode2 min read

Lite mode: the same open models at up to 50% off

We're very excited to introduce Lite mode: the same models from Keln at up to 50% off. You can turn it on by setting service_tier to lite.

Since 28 October, some models on Keln have 2 modes. Until now every request ran in Standard, and today we're introducing Lite. Lite runs the same model at up to 50% off the Standard price per token. You can turn it on with one field in the request, or once per key in the portal.

What stays the same in Lite

A Lite request runs the same model as a Standard one. Keln serves it at the Lite price and runs the same quality verification as in Standard, so a Lite request gets the full model. Reasoning levels, tool calls, prompt caching and zero data retention work the same way in both modes.

Lite has its own price for input, cached input and output, and all 3 are lower than the Standard price. On Models & Pricing, each model with Lite shows its discount in a green tag, Lite -35%, and lists its Lite price.

Speed targets

On average, a Lite request runs a little slower than a Standard one, though both modes meet the same targets: a first token within 2 seconds and at least 50 tokens per second.

How to turn it on

Add "service_tier": "lite" to the request, or set a key's Mode to Lite in the portal.

curl https://api.keln.ai/v1/chat/completions \
  -H "Authorization: Bearer $KELN_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-oss-120b",
    "service_tier": "lite",
    "messages": [{ "role": "user", "content": "Hi" }]
  }'

The Lite docs show the setup for each API and the response headers.

When Lite can't take a request

Lite is best effort. If Keln can't serve a request at the Lite price, it runs the request in Standard mode at the Standard price. Each response says which mode served it, and you pay that mode's price. The Usage page shows how much Lite saved you.

When to use Lite

Lite is best for work where the price per token matters more than top speed, such as batch jobs and evals. A chat product where users wait for each reply may be better on Standard, since Standard is a little faster on average.

LitePricingAPI

Try Lite mode

Create a key, add credit, and set service_tier to lite or set the key's Mode to Lite.