Developer Simon Willison released version 0.33 of his open-source command-line tool and Python framework llm on August 22, 2026. The tool serves developers and researchers as a unified interface to interact with diverse language models locally or via cloud APIs directly from the terminal. The update introduces foundational architectural improvements along with new functional capabilities tailored to recent reasoning models.
Under the hood, llm 0.33 migrates to version 3.x of the official OpenAI Python library while updating its underlying HTTP dependency to httpx2. These technical changes improve network stability during long-running streaming sessions and ensure compatibility with current API specifications across multiple model providers.
A major workflow enhancement in this release is support for template chaining. Developers can now supply the -t or --template flag multiple times in sequence to combine modular configurations. This allows users to decouple model-level settings, such as reasoning intensity, from specific prompt templates without creating redundant configuration files for every permutation.
The update also introduces native support for the reasoning_summary parameter, offering auto, concise, and detailed options for compatible reasoning APIs. For vector embeddings, version 0.33 adds the --key parameter to both llm embed and llm embed-multi commands, enabling isolated key injection without modifying global environment states or plugin configurations.
Alongside the core utility, the llm-gemini plugin received an update to version 0.33. This companion extension adds support for reasoning traces and server-side code execution. These updates align the command-line ecosystem with current industry shifts toward transparent model thinking processes and integrated execution environments.