Gemma variants (the first-party derivative family)
"gemma-4" names a lineup of general models. Gemma names something larger: a base that google repeatedly re-trains for one domain and ships under its own name. The awesome-gemma catalog lists 16 such variants, more entries than the core models get.
The catalog's list (2026-08-20)
| Variant | What it is |
|---|---|
| MedGemma | medical text and image comprehension |
| TxGemma | therapeutic-development research |
| PaliGemma 2 | vision-language, detailed image understanding |
| ShieldGemma 2 | image-safety classifier (built on Gemma 3) |
| EmbeddingGemma | compact embeddings for retrieval and on-device use |
| FunctionGemma | base for specialized function-calling models (wire format: functiongemma-formatting-and-best-practices) |
| TranslateGemma | translation across 55 languages |
| DiffusionGemma | experimental discrete-diffusion text generation on Gemma 4 |
| T5Gemma 2 | encoder-decoder |
| RecurrentGemma | the recurrent Griffin architecture |
| VaultGemma | trained with differential privacy (1B) |
| DataGemma | grounding answers in Data Commons |
| Gemma-APS | abstractive proposition segmentation |
| Gemma Scope 2 | sparse autoencoders and interpretability tooling for Gemma 3 |
| DolphinGemma | dolphin audio, for researchers studying their communication |
| Cell2Sentence-Scale | single-cell biology; Gemma 2 27B, fine-tuned by the van Dijk lab |
Descriptions are the catalog's, one line each — it evaluates none of them. All but the last are
published under Google's own Hugging Face org; Cell2Sentence-Scale is third-party
(vandijklab), which is why it sits at the edge of the list rather than in it.
What the shape says
Generations run in parallel, not in sequence. ShieldGemma 2 and Gemma Scope 2 are built on Gemma 3, Cell2Sentence on Gemma 2, DiffusionGemma on Gemma 4. A variant is not re-based every time the core family ships, so "Gemma" in 2026 spans at least three live generations at once. Anyone reading a variant's capabilities off gemma-4 will be wrong about most of them.
The architecture is not fixed either. T5Gemma is encoder-decoder, RecurrentGemma is Griffin, DiffusionGemma is discrete diffusion — three departures from the decoder-only transformer that gemma-4 describes. The shared thing is the name and the weights lineage, not the design.
Two of these are research instruments rather than products. Gemma Scope 2 ships sparse autoencoders for reading Gemma's internals, and DolphinGemma exists to study animal communication. An open-weight family is being used as laboratory equipment, which is a use no pricing or benchmark comparison in this wiki captures.
First-party derivatives are the mirror of the third-party ones. gemma-4-31b-it-scotoma showed outsiders editing Gemma's safety layer under Apache-2.0. This page shows the licensor doing the same specialization deliberately, and getting distribution the outsiders do not — official HF collections, a Kaggle impact hackathon, docs. Same legal freedom, very different reach.
Caveats
Sourced from a T3 catalog (awesome-gemma) that relays Google's collection pages. Parameter counts, licenses and release dates for individual variants are not recorded here because the catalog does not give them — only VaultGemma's 1B and Cell2Sentence's 27B appear in the source at all. Verify at the model card before any variant's specifics are used.
Related
gemma-4 · awesome-gemma · google · open-weight-models · gemma-4-technical-report · gemma-4-31b-it-scotoma