This week on the channel 3 videos this week

Maarten Grootendorst takes Gemma 4 apart, one layer at a time

Two DeepMind explainers on the open model's design, plus a voice AI conversation with Sierra, made up the week on Google for Developers.

Channel:

Google for DevelopersStory of September 29, 2026Trari Newsroom

Why this story? The channel's weekly summary: what went up this week, what is drawing views now, and where the channel stands among similar channels.

Gemma 4 12B: The encoder-free model explained

"Per-Layer Embeddings (PLE) in Gemma 4 explained" is the kind of video Google for Developers exists for: a short, focused lesson on one idea inside Google's open model family. In it, Google DeepMind's Maarten Grootendorst walks through how the smaller Gemma 4 variants, E2B and E4B, give each token its own embedding at every layer, a trick the description says lifts performance without adding to the parameters used during computation.

It was one of two Grootendorst explainers this week. The other, on Gemma 4 12B, covers a model the description calls encoder-free, taking in audio and images without separate encoders. It has the biggest total of the three uploads. The PLE video, only four days old, is the one gathering views fastest right now.

The third upload changed the voice, literally: Valeria Wu of Google DeepMind and Soham Ray of Sierra AI on real-time speech models, from response latency to switching languages mid-sentence. It opened briskly and has since settled.

The channel as a whole is running well above its usual daily pace. It sits third by subscribers among the technology companies around it, just ahead of Google Chrome, while Android Developers, its sibling channel, is also a little busier than normal.

Published on Trari