The latest Qt AI Assistant release adds custom coding instructions to every prompt, reducing the effort required to implement your own coding preferences or the project’s coding conventions. Furthermore, the 0.9.8 release also introduces support for Google’s Gemini 3 Pro Preview LLM and OpenAI’s GPT-5.1 LLM.
Custom Coding Instructions
You can now add their coding instructions to an agent.md file to align coding suggestions from LLMs with user-, project, or company-specific guidelines. Custom instructions are applied to natural language prompts and smart / commands. Custom instructions take priority over default coding instructions defined for smart / commands. You can set custom instructions on the project level in the Qt Creator IDE. You can create an agent.md instruction file using the Qt Creator file creation wizard, using a set of default coding instructions for Qt Quick in Qt 6 releases and for C++.
Gemini 3 Pro Preview Support
The latest Gemini version is scoring an 85% result on our 100 QML coding test tasks, making it the best QML coder in 2025. That is a 27% improvement over Gemini 2.5 Pro. It is the first time that Anthropic’s Sonnet LLMs have been beaten in two years of benchmarking (Claude Sonnet 4.5 scoring 75%).
The Gemini 3 Pro takes the latest Qt documentation into account. Its training data cut-off was in January 2025. It is the first LLM to give a meaningful response to a TreeView challenge. It creates shadows for Buttons and saturation for images using the MultiEffect visual effect type correctly. It still adds QtQuick.Controls and QtQuick.Layout imports too often. It also has problems with the rotation of the Dial's handle.
Considering the significantly lower price of $2 USD / million input tokens compared to $3 USD / million input tokens for Sonnet 4.5 API access (as of December 2025), it is an LLM that should be seriously considered by developers writing QML-based user interfaces.
GPT-5.1 Support
- GPT-5.1 replaces GPT-5. The chat-based coding performance of GPT-5.1 remains at the same level as GPT-5, achieving a 58% success rate on the QML100 benchmark. It adds frequently unnecessary import libraries. It has learned a few new things from the Qt 6 release, for example, implementing regular expression validation correctly with the RegularExpressionValidator QML type instead of the old RegExpValidator. It still creates visual effects based on QML types in Qt 5 releases.
Meanwhile... - Support for Llama 3.3 70B and DeepSeekCoder has been deprecated due to the limited usage and coding performance.
- Embedded QML linter upgraded for Qt 6.10 compliancy.
- Qt Quick Test cases are created now without repeating the application code, but linking to the application project.
- LLM configuration has been streamlined for a better first use experience.
