Fine-tune Gemma 4 on your data
Google's Gemma 4 models that read images and text, in dense sizes and one mixture of experts. Supported by the training and serving software Tensorant runs. A deployment served Gemma 4 E2B on Tensorant on 11 November 2026. No training run has been validated yet, so start with a trial.
- Status
- Supported
- Input
- Text and images
- Kind
- Dense and mixture of experts
- Trains with
- LoRA, QLoRA
- Suggested sizes
- LoRA from1 GPU, about 33 GB
Sizes
Sizes and GPUs
Each size with the GPUs one machine needs to train and to serve it. The figures are the estimates Tensorant checks a model against; a trial measures the real use. See GPU memory.
| Model | Parameters | Train with LoRA | Train with QLoRA | Serve | Status |
|---|---|---|---|---|---|
| google/gemma-4-E2B-itSmallest Gemma 4 | E2B | 1 GPUabout 16 GB | 1 GPUabout 7 GB | 1, 2 or 4 GPUsabout 14 GB in all | Supported |
| google/gemma-4-E4B-it | E4B | 1 GPUabout 23 GB | 1 GPUabout 9 GB | 1, 2 or 4 GPUsabout 21 GB in all | Supported |
| google/gemma-4-12B-itNewer, less proven in serving | 12B | 1 GPUabout 33 GB | 1 GPUabout 12 GB | 1, 2, 4 or 8 GPUsabout 31 GB in all | Supported |
| google/gemma-4-26B-A4B-itMixture of experts; an 80 GB GPU to train | 26B, 4B active | 1 GPUabout 67 GB | 1 GPUabout 61 GB | 1, 2, 4 or 8 GPUsabout 65 GB in all | Supported |
| google/gemma-4-31B-itAn 80 GB GPU to train | 31B | 1 GPUabout 80 GB | 1 GPUabout 24 GB | 1, 2, 4 or 8 GPUsabout 78 GB in all | Supported |
- Training needs the memory shown on each GPU: more GPUs on a machine train faster, but each holds the whole model.
- On a vision model, QLoRA trains with DPO only; SFT trains with LoRA.
- Serve: the numbers of GPUs of one machine the base model can be split across, and the memory it needs in all.
- Training on a GPU isn't confirmed yet for E2B and E4B. Serving the E2B base model was checked on Tensorant.
Training
How it trains
Train Gemma 4 with LoRA or QLoRA, by supervised fine-tuning (SFT) or DPO on preference pairs.
- SFT trains with LoRA, on a version with images or of text alone. On text, the vision tower stays frozen and the served model still reads images.
- DPO trains with LoRA or QLoRA, on one GPU.
- RL trains text models only.
- Use the instruction-tuned
-itcheckpoints: the base ones have no chat template. - The 26B A4B model trains adapters on its attention and shared layers, not on its experts.
Every training run, trials included, pays a training-token fee for the tokens it trains on, by the model's size. Images are on every organization.
Serving
How it serves
Publish a trained model under a name and call it through the OpenAI-compatible API, with the OpenAI SDK or any HTTP client. The name is what your application sends as model.
- A LoRA adapter you train serves on top of its base model, on the GPUs in the Serve column.
- The base model serves on its own too, in a test session or for a baseline. A deployment splits it across the GPUs of one machine: the Serve column lists the numbers each size allows.
- A served vision model reads images and text in the same chat request.
- The 12B model is newer and less proven in serving than the other sizes.
Steps
Fine-tune Gemma 4 on Tensorant
Create a project
Create a project with
google/gemma-4-12B-itas its base model, or pick another size from the list. The model check reads the model before anything starts.Import your data
Upload files, paste text or import a Hugging Face dataset as sources.
Review and freeze a version
Review the examples and approve the ones to train on, then freeze them into a version with training, validation and test splits.
Score the base model
Evaluate the base model on the version's test split. This baseline is what a training run is compared with, and a full run needs one.
Run a trial, then train
A trial trains a small sample to check the settings, the GPU memory and the speed. Then start the full run. LoRA on
google/gemma-4-12B-itneeds 1 GPU, about 33 GB.Compare and publish
Evaluate the trained model on the same test split and compare it with the baseline. Then publish it and call it through the OpenAI-compatible API.
Docs
Supported models
Every family's architecture, status and notes, and what a base model needs.
Model family
Gemma 3
Google's Gemma 3 models that read images and text, from ungated copies that need no access request.
Model family
Qwen3.8 Flash-Next
Qwen's large mixture-of-experts model that reads images and text, trained as an experiment on its text with LoRA on one machine of 8 NVIDIA B200 GPUs.
Start with one dataset.
Create an account and import your data, or book a demo and we will show you the product.