NEW GEMMA 4 beats GPT-5.4: The A4B Model

Google released yesterday 4 new OPEN AI Models. I test in this video the 2 most powerful versions of GEMMA 4. The dense GEMMA 4 31B and the GEMMA 4 MoE 26B-A4B model on pure scientific reasoning. Logic and causal reasoning. The results are surprising. Note: it is beneficial to let those tiny 4B models run longer. It may come up with a first solution after 1 minute, but for a better solution I let them run 10 minutes and more. As you see in my video, it took another run with optimize your reasoning for it to start exploring new alternatives. If you work locally, maybe your job is not time critical, and you want the best possible result. My advice: Give Gemma 4 A4B time to think. It will improve on its solutions. 00:00 GEMMA 4 Open Models 01:27 Live Test GEMMA 4 31B and A4B 10:17 First Validation run 31B and A4B 16:45 Second Validation run 31B and A4B 19:42 Optimize reasoning GEMMA 4 models 26:13 Verify result of GEMMA 4 A4B MoE 27:56 GEMMA 4 compared to other AI models #aitesting #airesearch #aiexplained #googledeepmind #gemma_4