Parameter Update: 2025-11

After the rush of releases in the past two weeks, this one felt almost calm by comparison. Nevertheless, there's still some cool stuff to dig into.

Google: Native image generation

In by far the coolest announcement of the week, Google finally gave us a first glance at multimodal image generation, first teased by OpenAI ten months ago. Launched in the most Google way imaginable, Gemini 2.0 Flash (Image Generation) Experimental is actually insanely cool, and if you haven't tried it, you should. It's still flawed in some major ways, and the abuse potential for this seems gigantic, but I, for one, am extremely excited about the possibilities this brings.

As a side note, it seems they also finally gave us "real" youtube link support (with actual video understanding, not just subtitles - even in the API!), so that's pretty cool too!

Google: Gemma 3

Finally merging the (slightly confusing) Gemma and PaliGemma lines of models, this week we got the third iteration, Gemma 3. Bringing both native image understanding and the best performance you can currently get on a single GPU, this seems like a really nice model to homelab if it weren't for the frankly insane license. What a shame, I was even ready to forgive them for lying about DeepSeeks VRAM requirements (no one is running R1 at bf16, Sundar).

Google: Gemini Robotics

More of a hype video than a release at this point, Google also brought us two new variants of Gemini designed for use on robots. They claim both generality and dynamic interactions in real time. With the wave of robotics announcements, that seems feasible, but I'm more excited to see their hardware at this point!

OpenAI: Agent Tools, Response API & literary model?

While no one was able to steal Google's thunder this week, I would be remiss not to mention OpenAI announcing some pretty cool new agent tools this week. In another (slightly less awkward) livestream, we got some cool new agent tools (Web Search, File Search, Computer Use), a new Responses API (basically a multimodal replacement for the Completions API that is basically an industry standard at this point) and Swarm being rebranded as Agents and moved into production. On Twitter, we got Sama showcasing some new model for creative writing, with people being very mad about it.

Baidu: ERNIE 4.5 & X1

While I was surprised to not hear anything from Baidu over the last little bit, it seems they've been busy building things to annoy OpenAI. While ERNIE 4.5 is a direct GPT-4.5 competitor, being touted for its "high EQ", X1 delivers the performance of "DeepSeek R1 at only half the price". With ERNIE 5 coming later this year, this does seem remarkably similar to OpenAI's communicated strategy. As for the models themselves, I haven't gotten to try them yet, so we'll have to take their word for it.