LLM 0.32 Separates Reasoning Traces From Normal Output

LLM 0.32 adds a practical split between what a model thinks through and what a program should actually consume.

The new release can display reasoning traces — the model’s interim reasoning text — separately on standard error. That keeps normal output clean for pipes, scripts and other command-line workflows, according to Simon Willison’s release notes.

For builders working with model output, that distinction matters. Reasoning traces can be useful when inspecting behavior, but they can also pollute the text a downstream tool expects to read. Sending them to standard error gives developers a way to watch what is happening without mixing diagnostic text into the main result.

Cleaner output for automated workflows

Command-line tools often rely on a simple contract: standard output carries the result, while standard error carries diagnostics. LLM 0.32 now fits that pattern more closely when reasoning traces are enabled.

The change means model reasoning can be inspected without breaking pipes, scripts or logs that depend on clean output.

That is especially useful when model responses are part of a larger chain. A script can continue reading the final answer from standard output, while a developer can review the reasoning trace separately when needed.

New server-side tools

The release also adds server-side tools for OpenAI code execution and web search. Those additions expand what LLM can call on during a run, while keeping the interface oriented around practical developer workflows.

LLM 0.32 also introduces a content-addressable SQLite message store. The goal is to reduce duplicate history in logs, making recorded conversations easier to manage and review.

Why it matters

The update is small in presentation but important in day-to-day use. Builders often need visibility into model behavior, yet they also need outputs that stay predictable when fed into other tools.

By separating reasoning traces from normal output, LLM 0.32 gives developers a cleaner way to debug model behavior without contaminating the text their scripts are built to handle.