Gemma model series

Google's Gemma line from Gemma 1 to Gemma 4: release months, sizes, MoE vs dense, context length, licenses, and which generation to use today.

What Gemma is

Gemma is Google's open-weight model line, published on Hugging Face under the google organization. It is the smaller, openly downloadable counterpart to Google's hosted models and has moved from text-only models to multimodal ones with audio, video and image input. The biggest practical change across the line is the license: Gemma 1 to 3n use Google's own Gemma license with an acceptance step, while Gemma 4 is Apache 2.0. Release months below come from Hugging Face repository creation dates.

Generations in order

Gemma 1 (February 2024). The first release, including the 7B model, created in February 2024 under the Gemma license. Access on Hugging Face is gated behind an agreement to Google's usage terms.

Gemma 2 (June 2024). The 27B instruction-tuned model dates from June 2024. It uses the same Gemma license and the same gated access.

Gemma 3 (March 2025). Four sizes: 1B, 4B, 12B and 27B. The 4B, 12B and 27B models have a 128K token context, the 1B has 32K, and maximum output is 8,192 tokens. The 27B card describes text and image input and over 140 languages. License: Gemma license. The 27B instruction-tuned repository was created in March 2025.

Gemma 3n (June 2025). Built for low-resource devices. Gemma 3n E4B has about 8B raw parameters but runs at an effective size of 4B through a MatFormer design. It accepts text, image, video and audio, with a 32K token context, under the Gemma license.

Gemma 4 (March 2026, with later additions). The first Apache 2.0 generation, with five models on the card:

  • E2B (2.3B effective) and E4B (4.5B effective, 8B total through per-layer embeddings): 128K token context, text, image, video and audio input.
  • 26B A4B: a mixture-of-experts model with 3.8B active parameters.
  • 31B: dense.
  • 12B Unified: 11.95B parameters, 256K token context, an encoder-free design that feeds image patches and audio straight into the language model. Its repository appeared in May 2026, after the first four.

The 26B A4B and 31B models have a 256K token context. The family includes a configurable thinking mode, native system prompts, function calling and 140 or more languages. Google also publishes quantization-aware (QAT) checkpoints, including 4-bit GGUF files, for the Gemma 4 sizes.

Which generation to use today

Gemma 4 is the newest generation in the google organization, and its sizes cover the ranges the earlier generations covered separately.

  • Default choice: Gemma 4. It is the newest, has the longest context in the line, and is Apache 2.0, so the license does not limit what you build on it.
  • Single GPU, strongest model: the dense 31B. For lower per-token cost, the 26B A4B activates only 3.8B parameters per token, though all its weights still need to fit in memory.
  • Mid-size with audio: the 12B Unified takes text, image and audio in one model.
  • Phones and small GPUs: Gemma 4 E2B and E4B, or Gemma 3n E4B if you need its MatFormer design.
  • Older Gemma work: Gemma 3 is the last generation under the Gemma license and remains a common fine-tuning base. If your lawyers already cleared that license, nothing forces a move, but new projects gain from Apache 2.0.

How to choose between sizes

The cards give enough to rank by what you need. For long documents, take the 12B, 26B A4B or 31B at 256K. For audio input, take E2B, E4B or 12B. For the smallest footprint, E2B has 2.3B effective parameters. For the highest quality in the family, Google's card reports the 31B at 85.2% on MMLU Pro and 89.2% on AIME 2026, which are vendor-reported numbers we have not reproduced.

Side lines

Google also publishes specialized Gemma branches: MedGemma for medical use, PaliGemma for vision-language tasks, and FunctionGemma for function calling. Each hub lists the models we track for that branch.

Gemma models can run on Aquanode GPUs.

Gemma generations

Every Gemma generation we track, oldest first. Each hub lists all of its models with VRAM at native, FP8 and INT4 precision.

HubModelsSizesSmallest native VRAMCheapest live fit for itEst. $/hr
Gemma 132.5B to 8.5B5.6 GBRTX 3070$0.088/hr
Gemma 2112.5B to 27.2B5.6 GBRTX 3070$0.088/hr
Gemma 311268M to 27.4B0.6 GBRTX 3070$0.088/hr
Gemma 3n45.4B to 8.4B12.2 GBRTX A4000$0.167/hr
Gemma 444420M to 32.7B0.9 GBRTX 3070$0.088/hr

Smallest native VRAM is the lowest requirement among the publisher's own checkpoints in the hub, at the precision they are published in; see the methodology.

Other Gemma lines

Side lines from Google: specialised or companion models outside the main Gemma generations.

HubModelsSizesSmallest native VRAMCheapest live fit for itEst. $/hr
MedGemma44.3B to 28.8B9.6 GBRTX A4000$0.167/hr
PaliGemma32.9B to 2.9B13.1 GBV100$0.110/hr
FunctionGemma1268M0.6 GBRTX 3070$0.088/hr

Run and fine-tune Gemma

Per-generation guides: engine support, chat template and context flags, then fine-tune memory for LoRA and QLoRA.

Gemma compared with other models

Side-by-side VRAM, context length and license for models of a similar size.

Best models by task

Gemma models appear on these ranked task pages.

Sources

Facts in the text above were read from these pages and are the publisher's own statements, not benchmarks run by Aquanode. Last reviewed 2026-10-07.

Start ComfyUI on a rented GPU.

© 2026 Aquanode. All rights reserved.

All trademarks, logos and brand names are the property of their respective owners.