Gemma 4 12B
Google's laptop-runnable open multimodal model with a unified encoder-free design.
Context window
256K tokens
Input / 1M tokens
Free
Output / 1M tokens
Free
Provider
Open-weight model under Apache 2.0 - free to download and self-host. Running costs depend on your own hardware (fits ~16GB VRAM at 8-bit quantization) or on inference provider pricing (e.g., Together AI, Ollama, Hugging Face Inference). · Data verified 2026-07-02
Gemma 4 12B is a member of Google's Gemma 4 open-model family. The Gemma 4 family launched April 2, 2026 with E2B, E4B, 26B MoE, and 31B dense variants; the 12B 'Unified' model was added mid-2026. It is a ~11.95B-parameter dense model with a 256K-token context window and a unified, encoder-free architecture that projects raw patches from text, image, and audio directly into a single decoder-only transformer, reducing multimodal latency. It is designed to run on consumer hardware (roughly 16GB VRAM at 8-bit) while approaching the quality of the larger 26B model, and is released under Apache 2.0.
Capability index
Relative estimates (0-100) to place this model against its peers, grounded in published benchmarks.
How to access it
Download the open weights from Hugging Face (google/gemma-4-12B or google/gemma-4-12B-it) or Kaggle and run locally via Transformers, Ollama, or vLLM. Also available through hosted inference providers.
Strengths
- ✓Open weights under permissive Apache 2.0 license
- ✓Native multimodality (text, image, audio) via a unified encoder-free design
- ✓Runs on consumer hardware (~16GB VRAM at 8-bit quantization)
- ✓Large 256K-token context window for an open model of this size
- ✓Multilingual (140+ languages) and supports function calling
Best for developers who...
When to choose it (and when not to)
Reach for Gemma 4 12B when...
- →You need an open, self-hostable multimodal model
- →Running on a single consumer GPU or a laptop
- →Data-privacy or on-prem requirements
- →You want a permissive license (Apache 2.0) for commercial use
Look elsewhere if...
- ✕You need the absolute highest quality (use the larger 26B/31B Gemma 4 or a frontier hosted model)
- ✕Complex multi-step reasoning where it falls behind the 26B variant
- ✕You prefer a fully managed API and don't want to run infrastructure
How to use it
- ›Use the instruction-tuned '-it' checkpoint for chat and assistant tasks
- ›Apply the official Gemma chat template so turn markers are formatted correctly
- ›Quantize to 4-8 bit to fit consumer GPUs with minimal quality loss
- ›Leverage native image/audio input rather than external preprocessing pipelines
Quickstart
Pythonfrom transformers import pipeline
pipe = pipeline("text-generation", model="google/gemma-4-12B-it", device_map="auto")
messages = [{"role": "user", "content": "Explain attention in transformers."}]
print(pipe(messages, max_new_tokens=256)[0]["generated_text"])Install `transformers` and `accelerate`. Accept the Gemma license on Hugging Face and authenticate with `huggingface-cli login`. Or run locally with `ollama run gemma4:12b`.
API model id: google/gemma-4-12B-it
Benchmarks
| Benchmark | Score | Notes |
|---|---|---|
| Total parameters | ~11.95B | Dense architecture; unified encoder-free multimodal |
| Context window | 256K tokens | Per official Hugging Face model card |