Techniquesmodel architecturemultimodal architecture

specific method · filed under model architecture

Configurable visual token budget

A configurable number of visual tokens used to represent images at variable resolutions.

sources: 3
model: 1
lab adopt it: 1
strongest: optional

How sources treat it

One count per evidence span, weakest treatment to strongest.

optional 3

Documented in

Evidence

3 spans quoted from the sources, strongest treatment first.

Gemma 4 supports variable image resolution through a configurable visual token budget, which controls how many tokens are used to represent an image.

optionalinference servingin Gemma 4Google DeepMind

google/gemma-4-26B-A4B-it · Hugging Face·Google DeepMind·2026-07-02

Gemma 4 supports variable image resolution through a configurable visual token budget, which controls how many tokens are used to represent an image.

optionalinference servingin Gemma 4Google DeepMind

google/gemma-4-31B-it · Hugging Face·Google DeepMind·2026-07-02

Gemma 4 supports variable image resolution through a configurable visual token budget, which controls how many tokens are used to represent an image.

optionalinference servingin Gemma 4Google DeepMind

Gemma 4 model card | Google AI for Developers·Google AI for Developers·2026-07-16

Filed alongside

Other methods under model architecture :: multimodal architecture.

Encoder-free multimodal architecture Early fusion multimodal training Discrete token encoding for audio Hierarchical patch encoder Mixed-modality training Joint multimodal decoding in a shared hidden space Native multimodal understanding 3×3 pixel-unshuffle downsampling Joint vision-language pre-training Raw audio projection into the LLM embedding space Vision Transformer (ViT)2×2 pixel-shuffle downsampling Audio encoder initialized from MiMo-Audio Audio Transformer (AuT)Causal streaming ConvNet codec decoder Data-parallel-first multimodal encoding Decoupled Encoder Process (DEP)Dedicated multimodal encoders Disaggregated encoder training Discarding the LLM after vision-encoder training Dual-format coordinate supervision Explicit text-string timestamps Four-frame audio patches with within-patch bidirectional self-attention From-scratch vision-encoder training with next-token prediction