Sikaru inference
Sikaru serves a model catalog behind OpenAI-shaped endpoints: chat completions,
completions, responses and models. Model IDs look like lab/model, for example
moonshotai/kimi-k3. Always take IDs from the catalog rather than from memory.
From the CLI
sikaru inference models # the catalog
sikaru inference models --search kimi --sort input
sikaru inference models MODEL # specs, pricing, reasoning efforts
sikaru inference chat MODEL "Explain a rainbow in one sentence."
sikaru inference chat MODEL "..." --stream # answer on stdout, reasoning on stderr
echo "..." | sikaru inference chat MODEL --system "Answer in plain words." -o json
chat also takes -t/--temperature, --max-tokens and --reasoning-effort. Read the
supported reasoning efforts from sikaru inference models MODEL before setting one.
Signed-in calls bill the saved project (sikaru config set project PROJECT).
From code
Use the inference client, Sikaru. It reads SIKARU_API_KEY from the environment.
The platform client for agents, runs and datasets is a different class,
SikaruApi, so don't mix them up.
from sikaru_sdk import Sikaru # AsyncSikaru for asyncio
client = Sikaru()
models = client.models.list()
reply = client.chat.completions.create(
model="moonshotai/kimi-k3",
messages=[{"role": "user", "content": "Explain a rainbow in one sentence."}],
)
print(reply.choices[0].message.content)
for chunk in client.chat.completions.create(model=..., messages=..., stream=True):
if chunk.choices and chunk.choices[0].delta:
print(chunk.choices[0].delta.content or "", end="")
import { Sikaru } from "@aadi-labs/sikaru-sdk/inference";
const client = new Sikaru({ apiKey: process.env.SIKARU_API_KEY });
const reply = await client.chat.completions.create({
model: "moonshotai/kimi-k3",
messages: [{ role: "user", content: "Explain a rainbow in one sentence." }],
});
Install with pip install sikaru-sdk or uv add sikaru-sdk, or npm i @aadi-labs/sikaru-sdk.
For reasoning output, pass include_reasoning. To move existing OpenAI-style code
over, swap the client and the model ID. Keep the key on the server.
Choosing a model
Compare candidates with sikaru inference models --sort input (or output) and read
each card. Recommend by the user's constraints: price, context length, reasoning
support. State the numbers you read; don't assert them from memory.