Gemma 4’s variants are named in a way that makes hardware planning harder than it should be. The E-variants advertise an effective parameter count, so E2B sounds like it needs 1.5GB and actually downloads at 7.2GB. The 26B-A4B advertises its active parameter count, so it sounds like a 16GB model and actually needs 18GB. Neither name is dishonest — they describe real architectural properties — but neither tells you what has to fit in your card. The sizes below do.
NVIDIA GeForce RTX 4090
24GB GDDR6X24GB VRAM is the minimum for Gemma 4 31B Dense at Q4. Nothing else in the consumer lineup comes close.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Quick answer by variant
- Gemma 4 E2B (~2B effective): 8GB+ VRAM at Q4, or 6GB on the QAT build — the name understates it
- Gemma 4 E4B (~4B effective): 12GB+ VRAM at Q4, or 8GB on the QAT build
- Gemma 4 12B Dense: RTX 3060 12GB and up — at ~7.6GB this is the variant a 16GB card is actually for
- Gemma 4 26B-A4B (MoE): RTX 3090 24GB used or RTX 4090 — 18GB of weights needs a 24GB card, and MoE means it still runs fast
- Gemma 4 31B Dense: RTX 4090 (24GB) minimum — RTX 5090 for Q5 or higher quantizations
VRAM requirements for every Gemma 4 variant
If you’re loading Gemma 4 through Ollama, see our Ollama VRAM Requirements guide for the per-quantization overhead Ollama adds on top of the raw model weights below.
| Variant | Architecture | Q4_K_M | QAT 4-bit | Min VRAM | Recommended |
|---|---|---|---|---|---|
| E2B (~2B effective) | Dense | ~7.2GB | ~4.3GB | 8GB | 12GB+ |
| E4B (~4B effective) | Dense | ~9.6GB | ~6.1GB | 12GB | 12GB+ |
| 12B Dense | Dense | ~7.6GB | ~7.2GB | 12GB | 16GB+ |
| 26B-A4B | MoE (4B active) | ~18GB | ~16GB | 24GB | 24GB+ |
| 31B Dense | Dense | ~20GB | ~19GB | 24GB | 24GB+ |
Every figure above is the published download size from Ollama’s Gemma 4 library, not an estimate derived from parameter counts — which is the whole point, since the parameter counts are what mislead here. Check them yourself before buying; Google reissues these builds and the sizes move.
The KV cache adds 2-4GB for longer conversations. That is what rules a 16GB card out of the 26B MoE: its Q4 weights are 18GB before any cache at all, and the 16GB QAT build leaves nothing. On the 31B Dense, the RTX 4090’s 24GB leaves only 4GB of headroom after model weights, which is enough for moderate context but not 16K+ conversations.
VRAM capacity memory bandwidth Specs are manufacturer figures. Bar lengths are scaled independently per metric.
Why the 26B-A4B MoE is the variant to watch
MoE (Mixture of Experts) means the model has 26B total parameters but only activates ~4B per forward pass. The result: you get 26B-class reasoning quality at 4B-class inference speed. On a used RTX 3090 (24GB) the modelled figure is ~55 tok/s at Q4 — faster than a traditional 14B dense model on a card two tiers below it.
The trade-off is memory, and it is bigger than the name suggests. All 26B parameters have to be resident regardless of which 4B activate, which is why gemma4:26b-a4b-it-q4_K_M is an 18GB download. That puts it on a 24GB card — a used RTX 3090 at ~$820 is the cheapest way in. What you get for that is 26B-class answers at 4B-class speed, which no dense model on the same card can match.
Performance benchmarks
Modelled from memory bandwidth at Q4_K_M, the way every throughput figure on this site is derived — see our methodology:
| GPU | E2B | E4B | 26B-A4B MoE | 31B Dense |
|---|---|---|---|---|
| RTX 5090 (32GB) | ~250 tok/s | ~180 tok/s | ~110 tok/s | ~42 tok/s |
| RTX 4090 (24GB) | ~170 tok/s | ~120 tok/s | ~80 tok/s | ~28 tok/s |
| RTX 5070 Ti (16GB) | ~130 tok/s | ~90 tok/s | Won’t fit | Won’t fit |
| RTX 4060 Ti 16GB | ~85 tok/s | ~60 tok/s | Won’t fit | Won’t fit |
| RTX 3090 (24GB, used) | ~110 tok/s | ~75 tok/s | ~55 tok/s | ~20 tok/s |
| RTX 3060 12GB (used) | ~45 tok/s | ~32 tok/s | Won’t fit | Won’t fit |
The 26B MoE’s speed advantage is clear — it runs faster than the 31B Dense on the same GPU despite similar total parameter counts. This is the MoE payoff in action.
GPU picks by budget
Around $250: E2B and E4B only
The RTX 3060 12GB (used, ~$250) handles both small variants with ease. 45 tok/s on the E2B is instant. If you only need lightweight Gemma 4, this is the floor — and in 2026 it is also the only card left under $300.
Check NVIDIA GeForce RTX 3060 12GB on Amazon→Buy on Shopee SG→$425-$1,050: the 16GB tier, and what it is really for
RTX 4060 Ti 16GB ($425) and RTX 5070 Ti ($1,050) both run the 12B Dense at ~7.6GB with plenty of room, and both run the E4B comfortably. Neither runs the 26B MoE: its Q4 weights are 18GB, and even the 16GB QAT build leaves nothing for KV cache.
Between the two, the gap widened this year. The 5070 Ti used to cost about $350 more; it now costs $625 more, for roughly 56% more speed on the same models. Unless you generate long-form output all day, the 4060 Ti 16GB is the better buy — and if the 26B MoE is what you came for, neither is the answer.
$2,200+: 31B Dense territory
The RTX 4090 ($2,200) is the entry point for the full 31B Dense model. At 28 tok/s with Q4, it is usable for interactive work. The RTX 5090 ($4,900) is the premium choice — 32GB VRAM means room for Q5 quantization and longer context windows, plus 42 tok/s makes it genuinely fast. Note the shape of this tier: you are paying roughly five times the 26B MoE’s entry price to move from a 26B-class answer to a 31B one.
NVIDIA GeForce RTX 5090
32GB GDDR732GB VRAM handles Gemma 4 31B at Q5 with room for context. The only consumer card comfortable with every Gemma 4 variant.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Common mistakes
- Buying a 16GB card for the 26B MoE. Its Q4 weights are 18GB. The “A4B” in the name is the active parameter count, which governs speed, not the memory footprint. This variant needs 24GB.
- Reading E2B and E4B as tiny models. They are named for effective parameters. E2B downloads at 7.2GB at Q4 and E4B at 9.6GB — the QAT builds (4.3GB and 6.1GB) are the ones that suit an 8GB card.
- Ignoring the 26B MoE entirely. Many buyers jump straight to the 31B Dense for “more parameters.” The 26B MoE delivers 90%+ of the quality at a fraction of the hardware cost. Test it first.
Final verdict
| Your goal | Best GPU | Price |
|---|---|---|
| E2B / E4B only | RTX 3060 12GB (used) | ~$250 |
| 12B Dense (best value) | RTX 4060 Ti 16GB | ~$425 |
| 26B MoE (budget) | RTX 3090 (used) | ~$820 |
| 26B MoE (best speed) | RTX 4090 | ~$2,200 |
| 31B Dense (budget) | RTX 3090 (used) | ~$820 |
| 31B Dense (best speed) | RTX 4090 | ~$2,200 |
| Every variant, max quality | RTX 5090 | ~$4,900 |
Our pick for most users: a used RTX 3090 at ~$820. It is the cheapest card that runs the 26B-A4B MoE — the standout Gemma 4 model — at ~55 tok/s with room for context. If the MoE is not what you are after, the RTX 4060 Ti 16GB at ~$425 handles the 12B Dense and both E-variants comfortably and saves you $395.
Check NVIDIA GeForce RTX 4060 Ti 16GB on Amazon→Buy on Shopee SG→For a detailed VRAM breakdown of every quantization level, see how much VRAM for Gemma 4. Upgrading from Gemma 3? Our Gemma 3 GPU guide covers the differences. And for the original Gemma family, check best GPU for Gemma. Budget-conscious buyers should also see our best budget GPU for local LLM roundup.