Overview
Gemma 4 launched in April 2026 with E2B, E4B, 26B A4B (mixture-of-experts) and 31B models; a 12B model was added later. All models read text and images, the small ones also take audio, and every size supports configurable thinking modes and native function calling. The “E” sizes use Per-Layer Embeddings to stay efficient on phones and laptops.
Highlights
- Sizes from E2B to 31B, including a 26B MoE with about 3.8B active parameters
- Thinking modes and native system-prompt support
- Audio input on E2B, E4B and 12B
- 128K context on small models, 256K on medium models
- First Gemma release under the Apache 2.0 license
Typical uses
On-device assistants, local coding and agents, multimodal apps and fine-tuning.
Things to consider
- Video and audio support differs by size; check the model card.
- Quantization-aware (QAT) checkpoints are the better choice for low-memory devices.
- Download sizes refer to the default quantized versions in the Ollama library; full-precision weights are larger.
Specifications are taken from the model card on Hugging Face, the Ollama library and the provider’s announcement. Benchmark results depend on evaluation settings; see how AI benchmarks work.