Sign in, open My Keys, and create a personal key.
The model id on the Models page is the model parameter.
A personal key belongs to you and to one project you are a member of. The secret is shown once, when you create the key. Copy it then.
Yes. Ask for one under Request a key on My Keys, and say what it is for. The same request covers work outside a project. If it is approved, you create the key there. The secret is shown once, to you, and passes through nobody else.
Point a tool that speaks the OpenAI API at https://llm.apache.org and use your personal access token as the API key. Many tools need only two variables:
export OPENAI_BASE_URL=https://llm.apache.org/v1
export OPENAI_API_KEY=<your personal access token>
GET https://llm.apache.org/v1/models with that key lists the models it can reach.
Setup for each tool:
A small output budget on a reasoning model can return HTTP 200 with empty content, a finish_reason of length, and no error. The model spent the budget thinking. Some models reason unless told not to, and others do the reverse. Where the tool allows it, turn reasoning off:
"chat_template_kwargs": {"enable_thinking": false}
Prompt and output share the context window. A client whose output budget is larger than that window is rejected before the prompt is counted. Raising the client's context setting makes that worse. The Models page lists each window.
A stop on a round token count is the output budget. Raise it. A stop on a round amount of clock time is a timeout.
For each model: the id to use, whether it is available now, its context window, its licence, where its weights and training data come from, and who can see your prompts.
The Models page states that for each model. Private means every deployment runs on infrastructure the ASF controls. External means every request is sent to a third-party provider. Mixed means a request may take either path, and you cannot choose which.
qwen3-8b and qwen3-8b-bf16 are the same model. They run on different hardware, and one of the machines cannot load the compressed weights the other uses. Answers should be equivalent. Use whichever the Models page shows as available.