SourcesGoogle DeepMind
publisher · 2 sources
Google DeepMind
2 sources read from this publisher, newest first, with the methods each one describes.
google/gemma-4-26B-A4B-it · Hugging Face
first party release · model card · 2026-07-02 · 17 methods
Mixture of expertsHybrid Sliding-Window and Global AttentionSparse expert activationEncoder-free multimodal architecturePer-Layer Embeddings (PLE)Configurable visual token budgetCSAM filteringEvaluation without safety filtersFunction callingSensitive data filteringStandardized sampling configurationContent quality and safety filteringContinuous monitoringHuman evaluationInterleaved multimodal inputNative system prompt supportStep-by-step thinking mode
google/gemma-4-31B-it · Hugging Face
first party release · model card · 2026-07-02 · 15 methods
Mixture of expertsHybrid Sliding-Window and Global AttentionSparse expert activationEncoder-free multimodal architectureExcluding prior thinking from multi-turn historyPer-Layer Embeddings (PLE)Configurable visual token budgetCSAM filteringEvaluation without safety filtersSensitive data filteringStandardized sampling configurationProportional Rotary Position EmbeddingUnified Keys and Valuescontrol-token-enabled thinking modemultimodal prompt content ordering