Gemma 4 System Requirements: What You Need to Run It on PC, Mac, and Cloud

Gemma 4 is Google's most capable open model — and if you've been wondering whether your current machine can actually run it, this guide has the exact numbers you need.

Google released four versions of Gemma 4 in 2026, ranging from a small 2-billion-active-parameter model to a 31-billion-parameter dense variant. The gap in hardware requirements between them is massive, so picking the wrong one for your machine either leaves performance on the table or crashes your system.

Here's a complete breakdown by model, hardware type, and what you need to get started.

---

The Gemma 4 Model Lineup

Gemma 4 comes in four variants:

Model	Architecture	Active Params	Best For
E2B	MoE	2B	Low-end hardware, quick inference
E4B	MoE	4B	Balanced — best for most users
26B A4B	MoE (4B active)	4B	High quality, efficient at runtime
31B Dense	Dense transformer	31B	Maximum quality, high VRAM needed

The E2B and E4B labels mean "Effective 2B" and "Effective 4B" — these are Mixture-of-Experts models that only activate a fraction of their parameters per token, which is why they can run on far less hardware than their total size implies.

---

PC Requirements (NVIDIA / AMD GPU)

All VRAM figures below assume 4-bit quantization (Q4), which is the standard for running large models locally. Running in full BF16 precision roughly triples the VRAM requirement.

Gemma 4 E2B:

• Minimum VRAM: 3.2 GB (Q4) / 9.6 GB (BF16)
• Compatible cards: GTX 1060 6GB, GTX 1070, RTX 2060, RX 6600
• This model runs comfortably on mid-range cards from 2017 onwards

Gemma 4 E4B:

• Minimum VRAM: 5 GB (Q4) / 15 GB (BF16)
• Compatible cards: RTX 2060 Super, RTX 3060 12GB, RTX 4060
• Best model for most PC users — punches above its weight thanks to MoE architecture

Gemma 4 26B MoE (A4B):

• Minimum VRAM: 15.6 GB (Q4) / 48 GB (BF16)
• Compatible cards: RTX 3090, RTX 4080, RTX 4090 (24GB VRAM)
• Only 4B parameters activate per forward pass — inference is fast despite the large total size

Gemma 4 31B Dense:

• Minimum VRAM: 17.4 GB (Q4) / 58.3 GB (BF16)
• Compatible cards: RTX 3090 or 4090 at minimum (Q4 only)
• BF16 requires a multi-GPU setup (like 3× A100 80GB) — not realistic for home use
• If you're regularly running 31B, cloud inference via Ampere.sh is often more cost-effective than buying a $1,600 GPU

RAM (System Memory):

• For E2B / E4B: 16 GB RAM (8 GB minimum, but expect slowdowns)
• For 26B / 31B: 32 GB RAM recommended
• Fast storage (SSD NVMe) is important — models load from disk on startup

---

Mac Requirements (Apple Silicon)

Mac users have an advantage: Apple Silicon's unified memory architecture allows the CPU and GPU to share the same memory pool, which significantly lowers the effective hardware threshold compared to discrete GPU setups.

MacBook Air M1/M2 (8 GB):

• Can run E2B at Q4 — tight, expect slower token speeds
• E4B technically loads but will throttle; not recommended

MacBook Air M2 (16 GB) / MacBook Pro M3 (18 GB):

• Comfortably runs E2B and E4B
• E4B at Q4 is a good fit — solid output quality with reasonable speed
• This is the sweet spot for most Mac users

MacBook Pro M3 Max / Mac Mini M4 Pro (24 GB):

• Runs E4B comfortably, plus the 26B MoE at Q4
• The 26B MoE at Q4 requires ~15.6 GB — fits within 24 GB with room for the OS

Mac Studio / Mac Pro (32 GB+):

• Runs all models including 31B Dense at Q4 (needs ~17.4 GB)
• At 64 GB unified memory, even some BF16 configurations are possible
• This is the recommended setup for running 31B with fast token speed

---

Running Gemma 4: Step-by-Step with Ollama

The easiest way to run Gemma 4 locally is with Ollama — it handles downloading the model, quantization, and serving automatically.

Step 1 — Install Ollama Download from ollama.com. Available for macOS, Windows, and Linux. The installer takes about 2 minutes.

Step 2 — Pull the model

ollama pull gemma4:4b

Replace 4b with 2b, 27b, or latest depending on your hardware. Ollama automatically downloads the appropriate quantized version.

Step 3 — Start chatting

ollama run gemma4:4b

You're now running Gemma 4 fully offline. No API keys, no usage limits, no data leaving your machine.

Alternative: LM Studio If you'd prefer a graphical interface without the terminal, LM Studio is a solid choice. You can search for Gemma 4 models inside the app and run them with a chat UI. Works on both Mac and Windows — useful if our terminal guide is still on your to-do list.

---

When Local Isn't Enough: Cloud GPU Options

If you're working with the 31B model at full precision, or if you want to run inference on demand without keeping your machine powered on, cloud GPU rental is worth considering.

Ampere.sh is a budget-friendly GPU cloud built for AI workloads — you can spin up an A10G or A100 instance, run your model for a few hours, and shut it down when done. Pricing is significantly lower than AWS or Google Cloud for comparable specs.

For most people experimenting with Gemma 4, E4B locally is more than enough. But for production workloads or research with the 31B Dense, cloud inference is the pragmatic path.

---

Key Takeaways

• E2B runs on almost any GPU with 4+ GB VRAM or an 8 GB Mac
• E4B is the best pick for most users — 5 GB VRAM or 16 GB Mac is sufficient
• 26B MoE needs 16 GB VRAM or 24 GB unified memory (Mac Mini M4 Pro)
• 31B Dense needs a 24 GB GPU (RTX 4090) or 32 GB+ Mac Studio for practical use
• For larger models or full-precision work, cloud GPU via Ampere.sh beats buying hardware
• Ollama remains the fastest path to local Gemma 4 — pull a model, run it in one command

---

FAQ

What is the minimum GPU to run Gemma 4? A GPU with at least 4 GB VRAM can run Gemma 4 E2B at 4-bit quantization. The GTX 1060 6GB or newer is enough. For E4B, aim for 8 GB VRAM minimum (though 12 GB is more comfortable).

Can I run Gemma 4 on a MacBook Air? Yes. A MacBook Air M2 with 16 GB unified memory runs E4B (5 GB Q4) without issue. An 8 GB M1/M2 Air can run E2B but will be slower. For 26B MoE, you need at least 24 GB unified memory.

Do I need an internet connection to run Gemma 4? Only to download the model initially. Once pulled, Ollama runs completely offline — no API calls, no cloud dependency.

What's the difference between E4B and 31B? E4B is a Mixture-of-Experts model with 4 billion active parameters — fast, efficient, good quality. 31B Dense is a traditional transformer with all 31B parameters active every pass — much higher quality but requires far more memory and compute.

Can I run Gemma 4 on a gaming laptop? Yes, if it has a discrete GPU. An RTX 3060 laptop (6 GB VRAM) can run E2B. The RTX 4060 Laptop (8 GB) handles E4B. For 26B MoE, you need the RTX 4090 Mobile (16 GB).

What about CPU-only inference? It works but is very slow. E2B on a modern CPU (Ryzen 9 or M-series) is usable for testing — expect 2–5 tokens per second instead of 20–60 on GPU. For real productivity use, a dedicated GPU or Apple Silicon is strongly recommended.

Is Gemma 4 free to run locally? Yes. Gemma 4 is open-weight and free to use locally under Google's Gemma Terms of Use, which allows personal and commercial use. The models are available on Hugging Face and via Ollama at no cost.