The 12B ships as a multimodal class with no encoder loaded, and its sliding-window attention overflows Neuron's fused-attention SBUF. Three surgical fixes and it serves Paris at ~15 tok/s on one inf2.8xlarge.
The middle child of the series. The 12B is the first Gemma-4 that's neither a MatFormer "effective"
model (E2B/E4B) nor a giant (31B/26B) — but it has its own two surprises: it's packaged as a
**multimodal* model with an encoder that isn't there, and its attention overflows a Neuron hardware
buffer in a way the smaller models never did. This is the short, clean port that only needed three
fixes on top of the E4B recipe — once I found them.*
A field report on running Google's Gemma-4 on AWS Inferentia2: mixed attention heads, the vLLM / optimum-neuron / NxD dead-ends,…
A side-by-side of all five Gemma-4 variants on Inferentia2 — PLE, KV-sharing, MatFormer, mixed attention, and a 128-expert MoE —…
Porting Gemma-4 31B (dense) to AWS Inferentia2: the tensor-parallel recipe that worked at 12B collapses at 31B, NxD ModelBuilder…
An in-depth explainer to Gemma 4 12B; a unified, encoder-free multimodal model!
Google DeepMind releases Gemma 4 12B, an encoder-free multimodal model with native audio that runs on a laptop.
An overview of Gemma 4 12B, a model designed to bring high-performance multimodal intelligence directly to your laptop.
