Gemma 3 has been released to the general public this week. With performance rivaling Deepseek 671B at only 27 billion parameters. According to Google themselves and independent review, Gemma 3 is roughly as good as Gemini 1.5 Pro, which is still considered a flagship model by today's standards.

With even more Google news, the geeks at Deepmind cooked up version of Gemini built to run on physical robotic hardware, similar to the people at Figure AI with their AI robot. So I guess start cranking that metal music because we're gonna get Robo-Recall in real life any time soon now.

Tencent a few days ago released their "Hunyuan" AI video generation model and it has taken the open source AI community by storm, the model produces impressive video output without the crazy resource draw that some other video generators use.

Apple released the M3 Ultra Mac Studio with a whopping 512GB of usable universal RAM, with people on reddit getting the entire 671B Deepseek model running on the system by itself.

Despite being a smaller company, Reka AI released a 21B model called Reka Flash 3 which producing surprising results for its low parameter count. It isn't widely available quite yet but it is most definitely a company to keep an eye on.

Phi-4 Multimodal released this week with a version of the Phi model with image support. Despite it not being the forefront it shows Microsoft is still interested in AI and still making strides in their own in-house attempt.

That's it for this week!