In “From llama.vim to Qt Creator using AI” I showed how I converted the llama.vim source code to a Qt Creator code completion plugin.

Code completion is cool, but people would want to chat with the chat bots, right?

My rough idea was:

1. Get the markdown text from the model
2. Parse it with md4c and generate html
3. Display the html with litehtml

Here is the gpt-oss 20B response to this prompt:

write me a Qt markdown renderer, using md4c as parser library, and qlitehtml for rendering using litehtml as a html browser

llama-server does come with a web server that would provide a nice chat interface, with a way to store the conversations, export them, and so on.

This is part of llama.cpp/tools/server/webui and is implemented using TypeScript and React. I had no knowledge about these technologies. But AI does!

At first I’ve tried with Qwen3 Coder 30B and gpt-oss 20B to convert the whole webui project (4208 lines). But, as it turns out both models had issues with this task.

Qwen3 did recognize that the task was too big and offered a simplified version:

Because the original code is > 4000 lines, I do not rewrite every single line – that would be a huge undertaking and would not help you understand how the pieces map.

gpt-oss started generating a cpp file that had only headers:

#include <QApplication>

// ... after 230 headers that were in part repeating

#include <QSortFilterProxyModel>

I’ve stopped the chat.

Since these were the models that I had, I had to select less source files that I wanted to convert.

Less is more!

On Friday 8th of August 2025 I was a bit frustrated with the lack of AI assistant support with Qt Creator!

Previously I only used GitHub Copilot with Qt Creator, but I didn’t like the idea of the LLM not running locally, running somewhere in the cloud.

That week I tested gpt-oss, Codestral, and Qwen3 locally on my Mac Studio M2 Ultra, with the task of changing some conditional CMake code to use generator expressions. I remember Qwen3 doing the best from all of them.

The above LLMs were running just fine with llama-server, and after failing to use them with the existing Qt Creator AI offerings (1, 2), I had a look at the llama.cpp website and, found out about their coding assistants. llama.vim and llama.vscode, but that was it. No Qt Creator support

Then I noticed that llama.vim was implemented in one single file. Next, I thought, what if I ask Qwen3-Coder-30B-A3B something like:

given the vim plugin delimited by triple backicks:

[...]

can you create a similar Qt Creator plugin ?

To my surprise it started with:

Qt Creator Plugin for LLaMA Infill Completion

I’ll create a Qt Creator plugin that provides similar functionality to the vim plugin you’ve shared. This will include auto-completion, context gathering, and integration with llama.cpp server.

Plugin Structure

The plugin will consist of:

1. Core plugin class with initialization
2. Auto-completion system with FIM (Fill-In-the-Middle)
3. Context gathering and caching mechanisms
4. Integration with Qt Creator’s editor framework
5. Configuration UI

Let me implement this as a complete plugin:

In November 2022 while recovering from a Covid infection I found a Lenovo ThinkPad X13s on ebay.com for 600$.

It had 16GB of RAM, 512GB of SSD, and a Qualcomm 8cx Gen3 CPU. These stats were all better than my previous Samsung Galaxy Book Go 5G laptop. The 13.3” display would be a bit less than the 14” display that Samsung had. But in contrast had the ability to expand the storage with a M.2 SSD slot.

At that time Lenovo was asking ~2000$ for such a laptop. So I decided to buy it. With transport and customs was still cheap

ThinkPad X13s

Finally a nice Lenovo Windows Arm64 laptop! It had a solid build feeling, better than the Samsung Galaxy Book Go. The keyboard was as one would expect from the latest Lenovo ThinkPads. Felt as good as the ThinkPad A485 that I have had a while ago.

I had no issues with the display as I had the Samsung Galaxy Book Go. So no display switching in this post!

The 7-zip benchmark result for arm64 is:

$ 7z b

7-Zip 23.01 (arm64) : Copyright (c) 1999-2023 Igor Pavlov : 2023-06-20

Windows 10.0 22621
ARM64 0 D4B.0 cpus:8 128TB f:2804EB8C1004
LE

1T CPU Freq (MHz):  2992  2993  2986  2993  2974  2976  2992
4T CPU Freq (MHz): 349% 2538   384% 2794

RAM size:   15789 MB,  # CPU hardware threads:   8
RAM usage:   1779 MB,  # Benchmark threads:      8

                       Compressing  |                  Decompressing
Dict     Speed Usage    R/U Rating  |      Speed Usage    R/U Rating
         KiB/s     %   MIPS   MIPS  |      KiB/s     %   MIPS   MIPS

22:      42120   682   6012  40975  |     364424   687   4522  31075
23:      38934   674   5884  39670  |     365481   704   4491  31615
24:      36525   673   5834  39273  |     366770   730   4405  32181
25:      33273   666   5705  37991  |     345648   711   4328  30755
----------------------------------  | ------------------------------
Avr:     37713   674   5859  39477  |     360581   708   4437  31406
Tot:             691   5148  35442

Here is how it fares to the other Arm64 CPUs:

That’s quite an improvement from the Qualcomm 8cx Gen2 generation, but still not beating the Apple M1.

For the CMake presets feature presentation in Qt Creator 9 I needed a cross platform Windows and Linux screencast.

My Windows Arm64 laptop was the perfect platform for the use case of registering a CMake preset with a self built Qt.

Quickly I found out that I only had one option to run Qt Creator as a Linux application, and that’s via Ubuntu 22.04 running under Windows Subsystem For Linux GUI (WSLg).

That’s because there is no Virtual Box or VMware Player running on Windows Arm64.

Also I’ve tried using Hyper-V, which WSLg under the hood uses, but the Ubuntu 22.04 image wouldn’t boot.

WSL2 and WSLg was the way to go! I’ve installed Qt Creator via sudo apt install qtcreator and then started Qt Creator via the Windows shortcut named Qt Creator (Ubuntu-22.04)! This is a short cut for C:\Windows\System32\wslg.exe ~ -d Ubuntu-22.04 qtcreator

Qt Creator would look like this:

It’s very hard not to notice that the windows title bar looks a bit weird. The generic window icon, the mouse cursor that does strange theme changes, and the bit fat window borders!

My goal was to have Qt Creator 9 look similar on Windows 11 as a native Windows Arm64 application and as native Ubuntu 22.04 Linux application running with WSLg.

We all know the Arm CPU architecture from smartphones, which have a long battery life and passive cooling.

Apple has shown with the M1/2 laptops that you can have a laptop that kicks ass with an Arm CPU.

In November 2021 I had a look to see if there was something like that for Windows.

I found a handful of models. Microsoft Surface Pro X, Lenovo Flex 5G, Acer Spin 7 at prices between 1000 - 1500$, and lastly Samsung Galaxy Book Go 5G at 800$ (400$ on eBay).

Samsung Galaxy Book Go

In Germany I could only buy the Samsung Galaxy Book Go “European” LTE version. Which came with a Qualcomm Snapdragon (TM) 7c Gen 2 CPU, 4GB of RAM and 128GB SSD.

I bought a refurbished model for 300€.

I took 7zip and run 7z -b to benchmark for arm64, x86_64 and x86. The results are here, below you have the arm64 results:

C:\Program Files (Arm)\7-Zip>7z b

7-Zip 21.04 beta (arm64) : Copyright (c) 1999-2021 Igor Pavlov : 2021-11-02

Windows 10.0 22000
ARM64 0 805.D0E cpus:8 128T f:804EB8C1004
LE

1T CPU Freq (MHz):  1931  2549  2550  2493  2488  2549  2545
4T CPU Freq (MHz): 346% 1903   354% 1952

RAM size:    3659 MB,  # CPU hardware threads:   8
RAM usage:   1779 MB,  # Benchmark threads:      8

                       Compressing  |                  Decompressing
Dict     Speed Usage    R/U Rating  |      Speed Usage    R/U Rating
         KiB/s     %   MIPS   MIPS  |      KiB/s     %   MIPS   MIPS

22:      12741   678   1828  12395  |     194291   705   2349  16568
23:      12388   722   1749  12622  |     185160   687   2333  16017
24:      11760   729   1735  12645  |     187951   720   2290  16491
25:      11119   753   1686  12696  |     185261   739   2232  16484
----------------------------------  | ------------------------------
Avr:     12002   720   1750  12590  |     188166   713   2301  16390
Tot:             716   2025  14490

The numbers by themselves do not mean much, but let’s compare them with the Apple M1 results from 7-cpu.com:

That’s not that good, isn’t it? Also 4GB of RAM, and 128GB SSD with no means to upgrade, made the offering a bit uncool.

Another thing that I’ve noticed was the screen quality. It wasn’t at the same level as my previous Lenovo laptops. The colors would change depending how I moved my head.

Notebookcheck.net has an article named Samsung Galaxy Book Go in review: Silent office notebook (archive.org copy), and their conclusion:

A better display could have made the Samsung Galaxy Book Go a good and inexpensive notebook.

In my previous post Using GitHub Actions with C++ and CMake I have provided a GitHub Actions yaml configuration file for C++ projects using CMake.

Building a project on GitHub Actions means always a build from scratch, for any given change, big or small. This takes time and wastes resources unnecessarily.

GitHub provides a way of caching dependencies to speed up workflows. The total size of cached files per repository is 2 GiB.

By having a look at the examples of various programming languages we can see that this is meant to cache package manager dependencies e.g. pip for python, npn for node, or gradle for java.

But, as it turns out, the caching mechanism can be used to cache compilation artifacts.

In this post I am going to provide a GitHub Actions configuration yaml file for C++ projects using CMake.

GitHub Actions is a CI/CD infrastructure provided by GitHub. GitHub Actions currently offers the following virtual machines (runners):

Each virtual machine has the same hardware resources available:

• 2-core CPU
• 7 GB of RAM memory
• 14 GB of SSD disk space

Each job in a workflow can run for up to 6 hours of execution time.

Unfortunately when I enabled GitHub Actions on a C++ project I was presented with this workflow:

./configure
make
make check
make distcheck

This is not something you can use with CMake though

Hello World

I am going to build the following C++ hello world program:

#include <iostream>

int main()
{
  std::cout << "Hello world\n";
}

With the following CMake project:

cmake_minimum_required(VERSION 3.16)

project(main)

add_executable(main main.cpp)

install(TARGETS main)

enable_testing()
add_test(NAME main COMMAND main)

TL;DR see the project on GitHub.

Coming from other build systems to CMake one will quickly learn that CMake can build only one configuration at a time. In practice you need to set up multiple build directories and configure/build with CMake for each and every one.

Autotools can do static and shared builds of libraries. For CMake most of the project would do a static build, then a shared build by setting the CMake variable BUILD_SHARED_LIBS to ON.

QMake can do debug and release builds at the same time, and as we can read at Qt for Android better than ever before, it can configure multiple Android architecture configurations at the same time.

What can we do to get the same level of convenience with CMake?

In this article I’m going to talk about building a C++ library with CMake, but it won’t be a CMake tutorial.

Let’s say you have a C++ library which depends upon a few open source libraries, which have a CMake project structure, but not necessarily done by the book (which means that they get only get built, and not deployed / installed)

Your library will include tests (unit-tests / integration tests), and the deployment can be just packing the headers and the binaries together in a tar.gz file.

This is not necessarily by the book, but it will do the job, and it could fit into any build system that the client has.

A book that one can use to do CMake right is Profesional CMake. Awesome CMake also has a great list of resources regarding CMake.

Coming back to the C++ library, which decisions do we take to build it? Shared library, static library, both?

In this article I am revisting an article from three years ago: “Speeding up libclang on Windows”, in which I was having a look at how the experimental Clang Code Model was handling a particular source code file.

With the help of Profile Guided Optimization I was able to go down from 10 seconds to 6 seconds.

In the meantime the Clang Code Model has been enabled by default in Qt Creator 4.7.

Three years ago I tested Qt Creator 3.6.0, Qt 5.5.1, LLVM/Clang 3.6.2, MinGW GCC 5.3.0, Visual C++ 2013/5. I tested on a Lenovo W510 Thinkpad with an “Intel(R) Core (TM) i7 CPU M 620 @ 2.67 GHz” CPU.

Now I am going to test Qt Creator 4.8.2, Qt 5.12.2, LLVM/Clang 7.0.1, MinGW GCC 7.3.0, and Visual C++ 2017. I upgraded my laptop to a Lenovo A485 Thinkpad with an “AMD Ryzen 7 Pro 2700U w/ Radeon Vega Mobile Gfx 2.20 GHz” CPU.

How many seconds would it take libclang to parse the file? TL;DR? 3 seconds!