- Published on
Llm 0.32rc2
- Authors
- Name
- DeepHorus
Simo Levy, the lead developer of the open-source command-line tool LLM, announced the release of version 0.32rc2 on 30th July 2026, hot on the heels of the first release candidate. This update not only fixes a critical dependency issue but also introduces two significant features: a new default model and a versatile CLI command for interacting with any OpenAI-compatible endpoint. The announcement, made via the project’s official blog, positions this release as a major step forward for developers who rely on LLM for prompt management and model experimentation.
The dependency fix resolves a packaging conflict that affected users upgrading from RC1, ensuring smoother installations across Python environments. More strikingly, the default model for users without a custom default has shifted from GPT-4o mini to GPT-5.6 Luna. This change reflects the rapid evolution of AI models, with Luna offering substantially better performance and reasoning capabilities. However, the upgrade comes at a slightly higher cost: $0.20 per million input tokens and $1.20 per million output tokens, compared to $0.15 and $0.60 for GPT-4o mini. Users can revert to the older model or choose the even cheaper GPT-5 nano ($0.05/$0.40) via simple commands.
The second headline feature is the new llm openai endpoint command, designed for running prompts, chats, and model listings against arbitrary OpenAI-compatible endpoints without prior configuration. This is particularly useful for developers testing local models via LM Studio or other proxy servers, as it eliminates the need to set up custom model configurations. Notably, these calls are not logged, which provides an additional layer of privacy for experimentation. The command even supports tools, as demonstrated in a one-liner example that queries a local Gemma model for the current LLM version and time.
The impact of this release is broad. For existing users, the default model change could alter costs and output quality without any action, potentially surprising those who track token expenses. The new endpoint command, however, is a game-changer for MLOps engineers and hobbyists alike, offering a lightweight way to benchmark and compare models from different providers. By using uvx --pre, users can even run LLM without installing it, making the tool more accessible than ever.
Industry reactions have been generally positive, with early adopters praising the flexibility of the new endpoint command on Hacker News. One developer noted, 'This kills the need for a dozen different CLI wrappers for local inference servers.' Others, however, expressed concerns about the default model cost increase, though the availability of cheaper alternatives eased those worries. The LLM team has not yet announced a final release date for 0.32, but the RC2 suggests it is imminent.
Looking ahead, this release signals a growing trend of tool consolidation in the AI ecosystem. As OpenAI-compatible endpoints become the de facto standard for model serving, commands like llm openai endpoint reduce friction for developers. The default model switch also indicates a broader shift toward more powerful, reasoning-focused models. With LLM’s continued evolution, it remains a key utility in the modern AI developer’s toolkit, bridging the gap between experimental models and production workflows.
