Simon Willison Refactors LLM Library to Support Frontier Reasoning and Tools

Simon WillisonSimon Willison

Simon Willison released an alpha version of his LLM Python library and CLI tool that moves beyond simple text prompts to support complex message sequences. The update introduces a streaming architecture for handling reasoning tokens and tool calls to maintain compatibility with frontier model capabilities.

Simon Willison, the creator of Datasette, released version 0.32a0 of his LLM Python library and CLI tool. The update replaces text-only prompts with a message-based system and a streaming architecture designed to handle reasoning tokens (internal model thinking and logic steps) and tool calls.

As frontier models shift toward agentic workflows, they no longer return simple strings. Models like Claude now produce internal reasoning and structured tool requests alongside standard text. This refactor ensures developers can capture these distinct typed parts without the library's core logic breaking or conflating different output streams.

You can now use the messages=[] array to pass conversation histories to models like GPT-5.5. The CLI highlights reasoning tokens in a different color and provides a --no-reasoning flag to suppress them. The alpha is available now via the LLM GitHub repository for testing.

View the full update on simonwillison.net
Simon Willison
Simon Willison
@simonw
X

I released LLM 0.32a0 this morning, a major backwards-compatible refactor of my LLM Python library and CLI tool for working with language models - the new changes should help LLM work better with reasoning models and other new frontier capabilities https://t.co/iLhtLrCQCL

3retweets70likes
View on X

Still wondering? A few quick answers below.

LLM is an open-source utility created by Simon Willison that provides a unified interface for interacting with thousands of different language models. It works as both a Python library and a command-line tool, using a plugin system to connect to proprietary APIs like OpenAI and Anthropic or local models running on a user's own hardware.

This alpha release introduces a major refactor that moves beyond simple text prompts to support a sequence of messages with specific roles like user and assistant. It also adds a new streaming architecture that can handle mixed content types, including text, reasoning tokens, and tool calls, allowing the library to better represent the outputs of modern frontier models.

The updated library can now distinguish between reasoning tokens, which represent a model's internal thinking process, and the final response text. In the command-line interface, these thinking tokens are displayed in a different color and sent to a separate output stream. Users can also suppress these tokens entirely using the new -R or --no-reasoning flag.

Yes, the 0.32a0 release is designed to be a backwards-compatible refactor. While it introduces new ways to interact with models using message arrays and event streams, the previous prompt-based methods still work. Under the hood, the library automatically converts single text prompts into the new message-based format to ensure existing code and workflows continue to function.

The alpha release introduces native serialization methods called to_dict and from_dict. These functions allow developers to convert a model response into a JSON-style dictionary that can be stored in any database or file system. This provides a flexible way to persist and reconstruct complex conversations without being forced to use the library's built-in SQLite logging system.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Share this update