The biggest model release this week was Google’s Gemma 4. It is Google’s best open model to date. The most impressive part about it is its ratio of parameter size to performance:
I’ll emphasize that the x-axis is on a log scale. Gemma 4 is absolutely tiny, yet ranks as the 3rd best open model on ArenaML, the popular chatbot arena.
These models being so small allows them to be feasibly run locally. If you have a Macbook from the past 2-3 years you can probably run 26b. For the 31b it is recommended to have 32GB of RAM which you will find on the higher end computers.
This is a huge unlock for anyone running their agents, like Hermes or Openclaw, locally. Maybe you don’t want the main session running through Gemma but you can easily have it for simpler dispatched tasks for absolutely no cost.
In other news OpenAI acquired the popular live streaming tech show TBPN, allegedly for hundreds of millions. The show only launched 18 months ago and regularly gets around 200k viewers.
This underscores something already widely thought: distribution and marketing have become increasingly important as advantages in tech even out. Whether or not this acquisition was because of the show itself or the talent it shows that brand is top of mind for OpenAI.
Apparently Anthropic just had one person doing marketing for much of the company’s history. With the recent PR blunders in Claude Code pricing, which I wrote about last week, I would not be surprised in a greater focus from them on their brand as well. And as the owner of this (tiny) newsletter and AI website I am not mad about it.
Thanks for following along,
Alex