krill bot

Optional local AI · Ollama + Qwen3 8B

The fish brought an 8B model.

Add local AI suggestions to Krill Bot for Krillion Puffer Party. The model runs on your computer. You still submit every answer.

Install the userscript first →

AI is an optional extra

Historical archive and local dictionary suggestions work without Ollama. Add a model if you want generated candidates for prompts where those sources are not enough.

You need an Ollama-supported computer with enough disk space and memory for Qwen3 8B. Generation speed and usable memory depend on your hardware. A GPU can help, but this project does not establish a minimum hardware specification or promise answers within a round.

Start with the basic helper. The Krillion userscript installation is separate from this optional model setup.

Set up Ollama and Qwen3

  1. Install Ollama using its official download and setup instructions for your operating system. Use the local model; Krill Bot does not require an Ollama cloud account or API key.
  2. Make sure Ollama is running, then download the configured model:
ollama pull qwen3:8b

If your installation does not already run Ollama in the background, start the local service in another terminal:

ollama serve

If it reports that the address is already in use, Ollama may already be running. Check its local model list:

curl http://127.0.0.1:11434/api/tags

The response should list qwen3:8b after the download completes. The userscript’s default endpoint is http://127.0.0.1:11434/api/generate. Use Ollama on the same computer as your browser.

Warm the brain cell

  1. Fully reload Krillion after installing the userscript, then open Puffer Party.
  2. Choose Warm 8B in the Krill Bot panel to load the configured model before you need it.
  3. The script requests generated suggestions when needed. Read them before choosing one; selecting an answer only fills the input.
  4. Use Regenerate to skip the displayed suggestions and request alternatives through the helper’s candidate selection.
  5. Choose Unload 8B when you want Ollama to release the model from memory.

Warming reduces the need to load the model during a round, but it does not guarantee low latency. The configured generation timeout is 45 seconds, so a slow request can outlast the time available to answer.

What the model sees

When generation is needed, the script sends category and candidate text to the local endpoint. It requests structured JSON with thinking disabled. No cloud model is required by this configuration.

The game remains the authority on acceptance and score. Qwen can invent answers, miss a restriction, or give a familiar answer. Generated candidates are not verified facts or guaranteed high-scoring answers.

The model name and endpoint are defined in CFG near the top of the userscript. The distributed script grants network access to 127.0.0.1 for Ollama. Changing the endpoint requires reviewing those permissions as well; the documented setup uses the default local endpoint.

When generation is not working

Connection refused or unavailable

Confirm Ollama is running on the same computer as the browser. Try the /api/tags command above. If that fails too, resolve the local service first using Ollama’s documentation.

Model not found

Run ollama list and check for qwen3:8b. Download it with ollama pull qwen3:8b if it is missing. The model name must match the script configuration.

The model is too slow or runs out of memory

Warm it before a round, close other memory-heavy applications, or use archive and dictionary suggestions without generation. There is no requirement to keep a local model loaded.

Ollama works, but the browser helper fails

Check the enabled script’s local-network permission and panel Diagnostics, then reload the game. The regression suite does not verify browser injection in a live session. Report your script, browser, and Violentmonkey versions in a redacted issue.

Need help with the browser panel?Userscript troubleshooting →