Simon Willison’s example and current Kimi docs list different reasoning_effort levels. Check the setting and preserve logs before an API comparison.
Simon Willison described his first Kimi K3.1 Pro API call on October 23 using the LLM command-line tool. He compared one visual task at different reasoning settings. For developers repeating the example, there is a practical discrepancy: his list of reasoning levels differs from the current API documentation. Check the setting and arrange to keep the output before starting a comparison.
As checked on October 26, Kimi’s guide lists low, high and max for reasoning_effort, with max as the default. Willison’s post names low, medium and max. That discrepancy does not establish that his earlier call was wrong. It does mean medium in the post should not automatically be treated as a currently documented setting. Record the exact model identifier, date and submitted parameter value for a new comparison.
Logging needs a separate check. The LLM documentation says llm openai endpoint neither registers the model nor logs the prompt and response in its normal history. Willison explicitly saved standard output and standard error to files. If you will need to review the result, keep those streams alongside the run settings rather than relying on terminal history alone. The files can contain the prompt and response themselves, so review them before sharing.
Service access and self-hosting are separate decisions. Moonshot’s announcement makes the API available while promising the full weights by November 4, 2026. A promised date does not establish that a particular weight file is already downloadable or will run on your hardware.
This is an analysis of a published experience and documentation, not a NODVIEW test of Kimi. We made no paid API calls. Willison’s single task does not establish performance on your code; preserving the prompt, settings and response gives you a basis for a reproducible comparison.