PewDiePie claims his fine-tuned AI model beats ChatGPT-4 and other leading tools

On February 26, 2026, popular YouTuber Felix “PewDiePie” Kjellberg uploaded a video with a bold claim. In a video titled “I Trained My Own AI… It beat ChatGPT,” he said the artificial intelligence model he trained performed better than ChatGPT-4 and several other well-known AI systems. Right at the start, he told viewers, “The deed is done. I can finally return to this channel ’cause I have done what I said I was going to do. I trained my own AI model.”

He then shared benchmark results that, according to him, backed up his statement. PewDiePie said, “I ran the benchmarks, official AI benchmark, and my model outperforms DeepSeek 2.5, way bigger model than mine. Facebook’s flagship model, LLaMA 4 Maverick, destroyed. Bang! Most importantly of all, my model outperforms ChatGPT’s 4 in like November or something.” The video quickly gained attention online, with fans debating how serious the comparison really was.

Felix “PewDiePie” Kjellberg explains how Qwen 32B fine-tuning and benchmark tests led to his ChatGPT claim

As the video moved forward, Felix “PewDiePie” Kjellberg made one thing clear. He did not build a brand-new AI system from zero. He explained, “I have not created my own AI. I have merely taken an AI model and trained it. It’s like stealing a child on the street instead of birthing one myself. It’s way more effective that way. Plus, it would cost millions and millions of dollars in infrastructure, which I do not have yet.”

He said his project was built on Qwen 32B, a model already strong at coding. However, he wanted it to perform better in a specific coding format.

He explained, “The model that I used was Qwen 32B, which is already amazing at coding, but I needed it to be amazing at coding in this format.” While testing, he focused on a benchmark called Ader Polyglot, which checks coding ability in six languages.

According to him, leading AI models did not perform as well as expected on that test. He said ChatGPT scored 18.2 percent, while Qwen 32B initially scored 8 percent.

The turning point came when he switched from the DIFF format to the HOL format. He explained that this change pushed the score to 16 percent.

To make it simple, he compared it to drawing a picture and adding a cloud without redrawing everything. His point was that format matters, and small technical adjustments can change results.

PewDiePie’s experiment adds to his growing interest in open-source AI. Last year, he also shared that he had hosted large AI models, including OpenAI’s GPT-OSS 120B, on his personal computer.

Whether his model truly “beat” ChatGPT in a broader sense is still open to debate. But his video clearly sparked fresh discussion about how AI tools are tested and compared.