Model · Google · Text
Gemma 4 (E2B / E4B / 12B / 26B-A4B)
The smaller Gemma 4 sizes under Apache 2.0 — E2B and E4B for phones and laptops, 12B and the 26B-A4B mixture-of-experts for a consumer graphics card. All read text, images and video; E2B, E4B and 12B also take audio. Thinking mode is built in.
- Modality
- Text
- License
- Apache 2.0 (Open)
- Parameter size
- E2B: 2.3B effective (5.1B with embeddings). E4B: 4.5B effective (8B with embeddings). 12B Unified: 11.95B. 26B A4B: 25.2B total, 3.8B active (MoE)
- Context window
- 128,000 tokens · up to 256k 128K tokens on E2B and E4B; 256K on 12B and 26B A4B, per Google's model card.
- Released
- March 31, 2026
- Last verified
- October 28, 2026
- Runs locally
- Yes
- Also handles
- vision, reasoning, speech-to-text
What you need to run it
- Mac
- Yes. Any Apple Silicon Mac (M1 or newer) with at least 8 GB (32 GB recommended) of memory. Newer chips are faster, but memory is what decides whether it runs.
- Windows / Linux PC
- NVIDIA works best. A graphics card with at least 6 GB (24 GB recommended) of video memory.
- No graphics card
- Works. Runs on a normal computer's processor, just slower.
The tiny E2B and E4B sizes run on an 8 GB or 16 GB Mac, even without a graphics card; the 26B-A4B needs a 32 GB Mac or 24 GB card because the whole model must fit in memory.
Numbers are for: 4-bit builds: E2B (smallest) to 26B-A4B (headline). Our estimate from the model’s size — the maker publishes no figure.
Strengths
- Apache 2.0 — no commercial restrictions
- E2B and E4B are built for phones and laptops; 12B and 26B A4B target a consumer graphics card or workstation
- 26B A4B activates only about 4B parameters per token, so it runs almost as fast as a 4B model
- Image and video understanding on every size; speech recognition and speech translation on E2B, E4B and 12B
- Switchable thinking mode for step-by-step reasoning
- Over 140 languages
Weaknesses
- Audio clips are capped at 30 seconds and video at about 60 seconds (one frame per second)
- No audio input on 26B A4B
- The small sizes trail the 31B flagship on long-context recall
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face (E4B) | weights | Apache 2.0; E2B, 12B and 26B-A4B in the same collection |
| Hugging Face (26B A4B) | weights | Apache 2.0 |
| Ollama | local | Quantised builds |
| Google AI Studio | hosted-api | Free tier |
Version history
- Gemma 4 31B Mar 2026
- Gemma 4 (E2B / E4B / 12B / 26B-A4B) Mar 2026 Current
- Gemma 2 9B Jun 2024 Deprecated
Official sources
- Model card (E4B) model
- Gemma releases docs
- Gemma 4 licence docs
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: estimate from size; Ollama gemma4 page; mlx-community gemma-4 builds.
- — Initial entry. Added 2026-10-28 from the popularity sweep: 7.43M Hugging Face downloads/month for gemma-4-26B-A4B-it, 4.03M for E4B and 3.06M for E2B. Google's release log dates E2B, E4B, 26B A4B and 31B to 2026-03-31 and the 12B Unified to 2026-06-03. Apache 2.0, 128K context (256K on 12B and 26B A4B), free to self-host. The 31B flagship has its own entry.
Esc