Model Request: Gemma 3n - by Alphabet/Google

[cougarten]

General Info: https://deepmind.google/models/gemma/gemma-3n/

Notable:

• might be best in weight class right now
• what they call "mix-and-match": you can downsize "model effort" on the fly without loading anything else in to memory
• different architecture to gemma3 (no clue if that's relevant to llama.cpp)
• can do vision

Weights:

• https://huggingface.co/collections/google/gemma-3n-preview-682ca41097a31e5ac804d57b

Terms:

• same as for any gemma https://ai.google.dev/gemma/terms

Ollama model request issue (>100 thumbs up):

• Gemma 3n ollama/ollama#10792

Can be tested on android phones with this apk (download in-app or hit [+] to add .task file):

• https://github.com/google-ai-edge/gallery/releases/tag/1.0.0

[The-Best-Codes]

Would love to see support now that Gemma 3n is stable:
https://huggingface.co/collections/google/gemma-3n-685065323f5984ef315c93f4

There are also community GGUFs:
https://huggingface.co/unsloth/gemma-3n-E4B-it-GGUF

[cougarten]

i guess it's done. at least downstream ollama has it now

[zhouwg]

FYI：huggingface/transformers#38300 (comment)

[zhouwg]

GGUF model: https://huggingface.co/ggml-org/gemma-3n-E4B-it-GGUF/tree/main
GGUF model: https://huggingface.co/ggml-org/gemma-3n-E2B-it-GGUF/tree/main

ngxson's PR in llama.cpp: #14400