Gemma 4

tl;dr: Gemma 4 is Google’s open multimodal family — five sizes from laptop-fit (E2B) to workstation-dense (31B), with text/image/audio on the small models and text/image on the larger ones. Thinking mode is on by default.

Variants

• gemma4:e2b — 2.3B effective (5.1B with embeddings). Text + image + audio. 128K context. Edge.
• gemma4:e4b — 4.5B effective (8B with embeddings). Text + image + audio. 128K context. Edge.
• gemma4:12b — 12B dense. Text + image. 256K context.
• gemma4:26b — 25.2B MoE (3.8B active, 8 of 128 experts). Text + image. 256K context.
• gemma4:31b — 30.7B dense. Text + image. 256K context. Frontier local.

gemma4 (no tag) routes to a sensible default per platform. MLX builds (*-mlx) and cloud variants (*-cloud) also exist on the library page.

Pull

ollama run gemma4:e2b
ollama run gemma4:31b
# frontier via Ollama's cloud:
ollama run gemma4:31b-cloud

What you get

• Native system / assistant / user chat roles and native function calling.
• Configurable thinking: prepend <|think|> to the system prompt to enable, remove to disable. Output uses <|channel|thought\n...<|channel|> for reasoning + answer.
• Variable-resolution image input: visual token budgets of 70, 140, 280, 560, 1120. Place images before text in the prompt.
• Audio input on E2B and E4B (CoVoST + FLEURS in their benchmark table).
• MLX builds optimized for Apple Silicon.

Hardware

Sampling defaults (from the library page)

• temperature=1.0, top_p=0.95, top_k=64. Reasoning models want the higher temperature.

Gotchas

• The library page does not state a license here — Gemma models are published under Google’s Gemma Terms upstream; check ai.google.dev/gemma before commercial deployment.
• Thinking tokens appear in the output by default. Don’t append prior thoughts to multi-turn history — the model expects only the final assistant turn.
• 26B is the MoE (A4B) — 3.8B active params, not the full 25.2B.
• Audio is only on E2B/E4B. The 31B dense has no audio input.

Related: Ollama for the runtime, llama.cpp for the engine, Open WebUI for a chat UI, LM Studio for a GUI runner, GPT-OSS as the OpenAI counterpart.

Untested here — facts from ollama.com/library/gemma4 fetched 2026-09-10.