From the model card
Fields below are copied from the tags and counters on the HuggingFace repository google/gemma-4-26B-A4B-it at our last fetch. They are set by the uploader, not verified by us; rows with no tag are omitted. How this page is made.
- Publisher (HF namespace)
- Pipeline tag
- image-text-to-text
- Library
- Transformers
- Weight formats
- safetensors
- License tag
apache-2.0— read the license file in the repo before relying on it- Lineage
-
- base model google/gemma-4-26B-A4B
- fine-tune of google/gemma-4-26B-A4B
- Papers cited
- arXiv:2607.02770
- Downloads (HF counter at last fetch)
- 7,837,690
- Likes (HF counter at last fetch)
- 1,476
- Model card
- https://huggingface.co/google/gemma-4-26B-A4B-it
Use cases
- Multimodal reasoning where per-token compute efficiency matters
- Local VLM deployment on infrastructure that cannot serve dense 30B+ models
- Image and text tasks requiring high model capacity at lower active parameter cost
- Research into MoE VLM architectures at open-weight scale
- Production VLM serving where throughput-per-GPU is a constraint
Pros
- Apache 2.0 license for commercial deployment
- MoE architecture reduces per-token active parameters vs. dense equivalent
- 26B total parameters provide strong multimodal capability
- Google DeepMind quality and HuggingFace Transformers native support
Cons
- MoE routing adds memory overhead — total weight footprint requires loading 26B parameters even with 4B active
- Load balancing across experts adds inference complexity
- MoE models can have expert load imbalance on specialized query types
- Newer Gemma generations may follow rapidly
- Quantized deployment of MoE models is more complex than dense models
Tags
transformerssafetensorsgemma4image-text-to-textconversationalarxiv:2607.02770base_model:google/gemma-4-26B-A4Bbase_model:finetune:google/gemma-4-26B-A4Blicense:apache-2.0eval-resultsendpoints_compatibledeploy:sagemakerdeploy:azureregion:us