Gemma 4

What is Gemma 4?

Gemma 4 is Google's open-weights language model with vision support. Released in 2026-05, it ships in 5 sizes (e2b, e4b, 12b, 26b, 31b) and is licensed Gemma License. The most popular use case is frontier-class multimodal with strong reasoning.

VRAM and hardware

The smallest e2b size needs at least 4 GB VRAM; the largest 31b needs around 8 GB.

Size	Min VRAM (Q4_K_M)	Recommended VRAM	Tokens/sec on 3090
e2b	4 GB	6 GB	~27 tok/s
e4b	4 GB	6 GB	~27 tok/s
12b	8 GB	12 GB	~22 tok/s
26b	17 GB	26 GB	~18 tok/s
31b	20 GB	30 GB	~17 tok/s

How to run Gemma 4 locally

Option 1: Ollama (simplest)

ollama pull gemma-4
ollama run gemma-4

Option 2: Mullama (production)

mullama pull gemma-4
mullama run gemma-4

Option 3: llama.cpp (CLI)

# Download a GGUF from Hugging Face (search "gemma-4 gguf")
./llama-cli -m gemma-4.Q4_K_M.gguf -p "Hello, AI!"

Option 4: Python with Mullama or llama-cpp-python

from mullama import Model, Context
model = Model.load("gemma-4.Q4_K_M.gguf", n_gpu_layers=99)
ctx = Context(model, n_ctx=4096)
print(ctx.generate("Hello, AI!", 256))

What you can build with Gemma 4

• frontier
• vision
• audio
• tools
• agentic

Hardware it fits on

• Apple Silicon (8GB): e2b in MLX
• Apple Silicon (16GB): e4b in MLX
• Apple Silicon (32GB): 12b in MLX or GGUF
• Apple Silicon (64GB+): 31b in MLX
• RTX 3090 (24GB): 12b at Q4_K_M
• RTX 4090 (24GB): same as 3090, ~30% faster
• RTX 5090 (32GB): 12b at Q8_0
• 2× RTX 6000 Ada (96GB total): 31b at Q4_K_M

Related models in the Gemma family

• Gemma 2 — 2B, 9B, 27B · Google-quality chat at small sizes
• Gemma 3 — 270M, 1B, 4B, 12B, 27B · single-GPU multimodal chat

Other models you might consider

• Llama 3.1 (Meta) — 8B, general-purpose chat, agentic workflows, code
• Llama 3.2 (Meta) — 1B, low-end hardware, mobile, edge deployment
• Llama 3.2 Vision (Meta) — 11B, vision + chat on desktop GPUs
• Llama 3.3 (Meta) — 70B, best-in-class 70B for desktops
• Llama 4 (Meta) — Scout 17B (109B MoE), vision + long context (10M tokens)
• Qwen 2.5 (Alibaba) — 0.5B, strongest 7B-32B on consumer hardware