Quick start

This page starts Spark, downloads a small model and sends a chat request. It uses SmolLM2 135M, a 145 MB model that is enough to see every step work.

Start Spark

SPARK_ADMIN_PASSWORD='choose-a-password' SPARK_STORAGE="$HOME/.curiosity/spark-data" spark

Spark listens on port 8080. Its database, models and KV snapshots go to ./storage unless SPARK_STORAGE names another folder.

Add a model

Open the console

Go to http://localhost:8080 and sign in as admin with the password you set.

Search Hugging Face

Open Discover and search for SmolLM2-135M-Instruct. Pick the repository bartowski/SmolLM2-135M-Instruct-GGUF.

Download a file

Choose SmolLM2-135M-Instruct-Q8_0.gguf and start the download. Downloads shows progress and speed. A dropped download resumes from where it ended.

See it in the library

The model appears under Models with the id smollm2-135m-instruct-q8_0.

The Models page listing a chat model and an embedding model, both loaded

You can also copy any .gguf file into $SPARK_STORAGE/models. It appears in the library on the next scan.

Create an API key

Open API keys, choose Create a key, and copy the key. It starts with sk-spark- and is shown once. Spark stores only its hash.

Send a request

curl http://localhost:8080/v1/chat/completions \
-H "Authorization: Bearer sk-spark-…" -H "Content-Type: application/json" \
-d '{"model": "smollm2-135m-instruct-q8_0", "messages": [{"role": "user", "content": "Hello!"}], "stream": true}'

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8080/v1", api_key="sk-spark-…")
reply = client.chat.completions.create(
model="smollm2-135m-instruct-q8_0",
messages=[{"role": "user", "content": "Hello!"}],
)
print(reply.choices[0].message.content)

import anthropic

client = anthropic.Anthropic(base_url="http://localhost:8080", api_key="sk-spark-…")
reply = client.messages.create(
model="smollm2-135m-instruct-q8_0",
max_tokens=256,
messages=[{"role": "user", "content": "Hello!"}],
)
print(reply.content[0].text)

The first request loads the model, which takes under a second for this one. Later requests reuse the loaded runner until the model has been idle for SPARK_IDLE_UNLOAD_MINUTES (30).

See what happened

Requests lists every request with its model, status, token counts and timings. Open one to see the prompt the chat template rendered, the answer and a timeline of each step.
Runners shows the runner process that holds the model, its memory and its log.
Playground sends a chat from the console, through the same endpoint.

The playground after one exchange, with tokens, speed and the share of the prompt served from the cache