This is a submission for the Gemma 4 Challenge: Write About Gemma 4

---

Which Gemma 4 Model Should You Actually Use? A Developer's Honest Guide

When Google DeepMind dropped Gemma 4 on April 2, 2026, the community response was immediate — 207,000 Ollama pulls in 48 hours, front page of Hacker News, and a same-day Ollama update to support all four variants. The hype was real. But so was the confusion.

Four models. Three naming conventions. Two architectures. One question every developer is quietly Googling:

Which one do I actually run?

This is that answer — practical, specific, with no benchmark-pasting.

---

First, Decode the Names

The naming is the first thing that trips people up. Let's fix that.

Model	What the name means	Architecture
E2B	Effective 2 Billion parameters	Dense + Per-Layer Embeddings
E4B	Effective 4 Billion parameters	Dense + Per-Layer Embeddings
26B A4B	26B total, 4B Active per token	Mixture of Experts (MoE)
31B	31 Billion parameters, all of them	Dense

The E in E2B and E4B stands for effective — not just a raw parameter count. These models use Per-Layer Embeddings (PLE), an architectural trick that lets them punch above their weight on constrained hardware. The A in 26B A4B stands for active — only 4 billion of those 26 billion parameters fire for any given token. That's the magic of Mixture of Experts.

If the names still feel weird, read them like this:

• E2B: "tiny but smart for its size"
• E4B: "the everyday laptop model"
• 26B A4B: "26B quality, 4B speed" ← the sleeper pick
• 31B: "no compromises"

---

The Hardware Reality

Before picking a model, be honest about your machine:

E2B — ~2–3 GB storage, runs on phones, Raspberry Pi, and anything with a CPU. If you're deploying to edge devices or need zero-latency local inference on minimal hardware, this is it. Don't use it for complex reasoning — it'll disappoint.

E4B — ~9.6 GB download via Ollama. This is the default ollama pull gemma4 variant for a reason. Runs comfortably on a 16 GB MacBook (M1 or later). Fast enough for interactive use. Good enough for most real tasks. If you're not sure which to pick, this is your answer.

26B A4B — The one most people overlook. You need around 24 GB of RAM (or a 24 GB GPU like an RTX 3090 or 4090). But what you get is near-31B quality at roughly E4B inference speed, because MoE only activates 3.8B parameters per token. Apple Silicon Mac with 32 GB unified memory? This is your best model.

31B Dense — 20 GB minimum RAM/VRAM, 24 GB recommended. Every single one of those 31 billion parameters fires for every token. No shortcuts. It currently sits at #3 among all open models globally on the Arena AI leaderboard. If you have a 4090 or an M2 Ultra, run this.

---

The Setup (Ollama, 5 Minutes)

Ollama is the fastest path from zero to running. Make sure you have Ollama 0.22 or newer — earlier versions don't handle Gemma 4 properly.

# Check your version
ollama --version

# Pull the model that matches your hardware
ollama pull gemma4:e2b    # phones, Pi, CPU-only machines
ollama pull gemma4        # E4B — 16 GB laptops (default)
ollama pull gemma4:26b    # 24 GB RAM — MoE, best quality/speed
ollama pull gemma4:31b    # 24 GB+ VRAM — maximum quality

# Run it
ollama run gemma4

One Critical Fix You Need to Make

Ollama's default context window for Gemma 4 is set to 4K tokens — but the actual models support 128K (E2B/E4B) and 256K (26B/31B). That default silently cripples long-context work. Fix it:

# Create a Modelfile with the right context
cat << 'EOF' > Modelfile
FROM gemma4
PARAMETER num_ctx 32768
EOF

# Build a custom named model
ollama create gemma4-32k -f Modelfile
ollama run gemma4-32k

For LM Studio users: search for Gemma 4 GGUF builds and use Q4_K_M quantization — it's the sweet spot between quality and RAM usage. Q5 if you have headroom to spare.

---

What Gemma 4 Actually Gets Right

Multimodal is native, not bolted on

Every Gemma 4 model handles text and images in a single model call — no separate vision pipeline, no switching endpoints. The E2B and E4B models go further and support audio input natively (up to 30 seconds), and the 26B/31B models handle video up to 60 seconds at 1fps. This isn't a demo feature. It's built into the base architecture.

128K context is usable in practice

A lot of models claim long context and then quietly degrade in quality past a few thousand tokens. Gemma 4 uses a hybrid attention mechanism — interleaving local sliding window attention with full global attention — specifically designed to maintain coherence at long range. For RAG pipelines, codebase analysis, or long-document work, this matters.

The license is actually open

Apache 2.0. Not Google's previous custom Gemma license. You can use it commercially, modify it, fine-tune it, and deploy it in products — no restrictions, no royalties. For developers building on top of a local model, this changes the calculus entirely.

---

The Decision Tree

Stop overthinking it. Use this:

What hardware do you have?
│
├─ Phone / Raspberry Pi / CPU-only → E2B
│
├─ 16 GB laptop (Mac, Windows, Linux) → E4B (ollama pull gemma4)
│
├─ 32 GB Apple Silicon or RTX 3090/4090 → 26B A4B ← don't skip this one
│
└─ 64 GB+ Mac or RTX 4090 and you need maximum quality → 31B Dense

What are you building?
│
├─ Mobile / edge app → E2B or E4B
│
├─ Local dev tool, coding assistant, RAG → E4B or 26B A4B
│
├─ Long-context document analysis, codebase reasoning → 26B or 31B (+ increase num_ctx)
│
└─ Fine-tuning for a specific domain → Start with 26B A4B

---

What This Actually Means

Here's the thing worth sitting with for a moment.

The 31B Dense model — the one that ranks third among all open models on Earth — runs on a consumer GPU. A single RTX 4090, the kind of card a serious gamer or developer might already own, is sufficient. No cluster. No cloud bill. No API rate limits. No data leaving your machine.

Two years ago, a model this capable required either a research institution's compute budget or a cloud provider's infrastructure. Today you pull it with one terminal command and it runs on hardware you might already own.

The E4B model — the second-smallest in the family — handles image input, supports 128K context, reasons in 140+ languages, and fits in 16 GB of RAM. That's a family phone or a mid-range MacBook.

Developers who internalize this shift will build very differently from those who don't. When inference is local and free, the calculus around what's worth building changes. Offline-first AI features stop being a niche edge case and start being a design choice. Privacy-sensitive applications that couldn't viably use cloud AI now have a real path.

That's what Gemma 4 is: not just a better model, but a different kind of constraint on what's possible.

---

Quick Reference

	E2B	E4B	26B A4B	31B Dense
Best for	Edge, mobile	Everyday dev	Quality + speed	Max quality
RAM needed	4 GB	8–16 GB	24 GB	20–24 GB+
Context	128K	128K	256K	256K
Multimodal	Text + Image + Audio	Text + Image + Audio	Text + Image + Video	Text + Image + Video
Ollama tag	gemma4:e2b	gemma4 (default)	gemma4:26b	gemma4:31b
License	Apache 2.0	Apache 2.0	Apache 2.0	Apache 2.0

---

Pick the model that matches your hardware. Fix the num_ctx default. Build something real.

That's it.