Kimi K3 in OpenHands

Coding agent that runs commands and browses · LiteLLM provider format · about 3 minutes

Before you start. Create an account (you get $2 to try), then open Token Management → Create token and copy the key. The key starts with sk-. Endpoints: OpenAI-compatible https://microquickjs.com/v1, Anthropic-compatible https://microquickjs.com. Model ID: Kimi-K3.

In the UI

Open Settings → LLM and switch on Advanced.
Custom Model: openai/Kimi-K3 (the openai/ prefix tells LiteLLM to use the OpenAI-compatible driver).
Base URL: https://microquickjs.com/v1
API Key: your ainetcafe key.
Save and start a new conversation.

In config.toml (headless or Docker)

[llm]
model = "openai/Kimi-K3"
base_url = "https://microquickjs.com/v1"
api_key = "sk-your-ainetcafe-key"
max_output_tokens = 32768
temperature = 0.6
native_tool_calling = true

Or as environment variables for the Docker image:

-e LLM_MODEL=openai/Kimi-K3 -e LLM_BASE_URL=https://microquickjs.com/v1 -e LLM_API_KEY=$AINETCAFE_API_KEY

Notes

Keep native_tool_calling on; K3 supports function calling natively and does worse with the prompt-based fallback.
OpenHands condenses history when it gets long; with K3's 1M window you can raise the condensation threshold.

Verify the key first

Thirty seconds now saves an hour of guessing later. Run this in a terminal; you should see a short reply and a usage block.

curl https://microquickjs.com/v1/chat/completions \
  -H "Authorization: Bearer $AINETCAFE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"Kimi-K3","messages":[{"role":"user","content":"Say OK"}],"max_tokens":5}'

If something goes wrong

Symptom	Cause	Fix
401 Invalid token	Key copied with a space, or key from another site	Re-copy from Token Management on microquickjs.com
403 quota / insufficient balance	Wallet at zero or plan quota used up	Top up in Wallet, or wait for the weekly plan reset
404 model not found	Model ID typed differently	Use exactly `Kimi-K3`
429	Rate limit for your tier	Reduce concurrency; limits rise with spend, or ask for a committed plan
Tool calls never happen	Tool calling disabled in the client, or thinking blocks stripped	Keep the client's default tool settings; K3 returns a `thinking` block before `tool_use`, do not filter it out

Still stuck? Email [email protected] with the request ID from the error body.