The v0.9.6 release adds support for GPT-5 and GPT-OSS to the Qt AI Assistant. GPT-5 increases the QML coding performance by 12% compared to GPT-4o. GPT-OSS 20B is the first royalty-free LLM for prompts that developers can host on their own computer.
GPT-5 Support
GPT-5 scores a 59% success rate on the QML100 benchmark for QML programming based on prompts. That performance puts it on the same level as Gemini 2.5 Pro. Not quite Claude Sonnet 4 territory, scoring a 77% on the QML100 benchmark, but still a significant improvement over GPT-4o. The training data for GPT-5 remains heavily focused on Qt5-compliant code, resulting in weaknesses in applying Qt6 syntax for QML types, such as RegularExpressionValidator, MessageDialog, and MultiEffect.
QML100 benchmark for prompt-based QML coding performance - September 2025
GPT-5 scores an 11% better result with a 73% performance than GPT-4o on the QML100FIM for QML code completion, but indentation and line breaks are a challenge for GPT-5. Therefore, we labelled GPT-5 as Experimental for code completion tasks.
QML100FIM benchmark for QML code completion performance - September 2025
GPT-OSS 20B Support
For QML code completions, Qt already provides fine-tuned models that can be deployed locally on the developer's machine (assuming they do have GPU acceleration), such as CodeLlama 13B QML. GPT-OSS 20B is the first royalty-free LLM that can possibly be deployed locally for expert advice following natural language requests/prompts.
GPT-OSS 20B delivers a reasonable QML code quality, similar to what GPT-4o does. The royalty-free GPT-OSS 20B scores a 46% success rate on the QML100 benchmark. Compared to the 47% that GPT-4o scored, and considering that it is a royalty-free model, that is a good performance showing the trend towards smaller language models. It is worthwhile mentioning that, in combination with the Qt AI Assistant and its embedded linter, developers can benefit from a coding performance of up to 54% in the QML100 benchmark when using GPT-OSS 20B.
GPT-OSS 20B does not "listen" well to instructions in the prompt. If there are multiple instructions,
then it may ignore some of them sporadically. We were not able to prompt it successfully for code
completion. Indentation and line break issues make the use of GPT-OSS 20B for code completion
inconvenient and not productive. Therefore, we decided not to provide support for GPT-OSS code
completion. Furthermore, due to the model's weakness in following complex instructions, we do not
recommend using GPT-OSS 20B for /qtest and /review tasks for Qt6-compliant code.
The training data of GPT-OSS 20B for QML programming is clearly outdated, very much based on Qt 5, even more outdated than for GPT-5. We also need to be aware that, while GPT-OSS 20B is a Mixture of Experts model that activates only a portion of its parameters during inference, the full model must still be loaded into memory on-device, requiring significantly more allocation than our CodeLlama 13B and 7B models. Most developers will find that they are unable to run GPT-OSS 20B with a satisfactory inference performance on their computers. It does work well, for example, on an Apple MacBook Pro M4 Pro with 48GB of memory, but only a few developers will have the latest and greatest yet.
In summary, GPT-OSS 20B is the first royalty-free, locally-running LLM for expert advice on QML programming. But with above mentioned limitations. It shows the way towards smaller, dedicated models, but its training data needs to be updated.
How to Install
You can install (or upgrade to) the Qt AI Assistant v0.9.6 from the Qt Creator Extensions view in Qt Creator 17.0.1 and later releases. Remember to activate the use of external repositories for extensions.
