By LLM Hardware team · Methodology ·
AI did the initial pull from Google's Gemma 3 model cards. VRAM figures were cross-checked against the methodology page formula.
Checked 2026-10-18. Author: LLM Hardware team. This site does not store or show Amazon prices. Check current price on Amazon.
Key fact: Gemma 3 27B at Q4 needs 18.2 GB. A 16 GB GPU does not fit it. Gemma 3 12B at Q4 needs 9.2 GB. An 8 GB GPU does not fit it.
How we got it: 27 × 0.5 × 1.2 + 2 = 18.2. 12 × 0.5 × 1.2 + 2 = 9.2. Ollama library on 11 October 2026 still listed gemma3:4b, gemma3:12b, and gemma3:27b.
Checked:
Key fact: Gemma 4 26B-A4B, counted at 25.2B total, needs 17.1 GB at Q4 and does not fit 16 GB. Gemma 4 31B dense, 30.7B, needs 20.4 GB and fits 24 GB.
How we got it: Ollama gemma4 library model table on 11 October 2026: 26B A4B is 25.2B total and 3.8B active; 31B dense is 30.7B. 25.2 × 0.5 × 1.2 + 2 = 17.1. 30.7 × 0.5 × 1.2 + 2 = 20.4. Tags present: gemma4:e2b, gemma4:e4b, gemma4:26b, gemma4:31b.
Checked:
Methodology — parameter math, quantization bytes, and the source list.
Gemma 4 is the current Ollama family. It does not replace this page's Gemma 3 commands. ollama run gemma4:26b is the 25.2B MoE, and ollama run gemma4:31b is the 30.7B dense model. Both need a 24 GB GPU under the formula above. ollama run gemma4:e4b counts 8B with embeddings, which is 6.8 GB at Q4 and fits 8 GB. Weights and the Ollama card are the source. In LM Studio, pick the Q4_K_M GGUF of the same size. The VRAM calculator uses the same formula. Sibling notes are on the Qwen3 hardware page and the Phi-4 hardware page.
Updated 24 October 2026 · Gemma 3 1B–27B · VRAM requirements · Consumer GPU guide · Ollama & LM Studio
Gemma 3, released by Google DeepMind in March 2025, is a family of open-weight models in 1B, 4B, 12B, and 27B sizes. The 27B model at Q4 is 18.2 GB, so it needs a 24 GB GPU. A 16 GB card does not fit it. The 12B model at Q4 is 9.2 GB and needs 12 GB, not 8 GB. All four sizes run on Ollama and LM Studio with no code required.
What is Gemma 3?
Gemma 3 is an open-weight model family from Google DeepMind. Released in March 2025, it comes in four sizes with instruction-tuned variants available for each.
VRAM is estimated with the site formula: parameters in billions × bytes per parameter × 1.2, plus 2 GB for the system. Q4_K_M is about 0.5 bytes per parameter, Q8 about 1.0 and FP16 about 2.0. Long context windows need extra headroom for the KV cache.
| Model | Params | Q4_K_M VRAM | Q8 VRAM | FP16 VRAM | Min GPU |
|---|---|---|---|---|---|
| Gemma 3 1B | 1B | 2.6 GB | 3.2 GB | 4.4 GB | 8 GB GPU |
| Gemma 3 4B | 4B | 4.4 GB | 6.8 GB | 11.6 GB | 8 GB GPU |
| Gemma 3 12B | 12B | 9.2 GB | 16.4 GB | 30.8 GB | 12 GB GPU |
| Gemma 3 27B | 27B | 18.2 GB | 34.4 GB | 66.8 GB | 24 GB GPU |
Gemma 3 27B at Q4 is 18.2 GB and does not fit a 16 GB GPU. Use the VRAM Calculator for context-length-adjusted estimates and KV cache projections.
Quantization trades a small amount of output quality for a large reduction in VRAM usage. For most users, Q4_K_M is the right default.
Q4_K_M — recommended
Cuts VRAM roughly in half versus FP16 with minimal quality loss on most tasks. The standard quantization for consumer hardware. All Ollama defaults use Q4_K_M. This is what to use unless you have excess VRAM.
Q8_0 — best quality
Q8_0 roughly doubles memory against Q4 and stays close to FP16 quality. Gemma 3 27B at Q8 comes to about 34.4 GB by our formula, which is more than a 32 GB RTX 5090 holds. Plan on a Mac with 48 GB or more of unified memory, and leave room for macOS and context. Use Q8 when you have that headroom.
FP16 — reference only
FP16 is full precision. Gemma 3 27B at FP16 comes to about 66.8 GB, more than any single consumer GPU holds. It is mainly for fine-tuning and reference runs, not daily use.
Q2/Q3 — small devices
Aggressive quantizations for very constrained hardware. Quality degrades noticeably — especially on reasoning tasks. Only consider Q2/Q3 if your device cannot fit Q4 and you need the model for lightweight tasks.
Find your GPU or Mac below. Each card shows which Gemma 3 models fit, and what does not.
RTX 4060 8GB
Runs:
Does not fit:
An 8 GB card fits gemma3:4b. It does not fit gemma3:12b or gemma3:27b. On 11 October 2026 the new buy box was not readable. Check current price on Amazon.
Intel Arc B580 12GB
Runs:
Does not fit:
12 GB is the comfortable tier for gemma3:12b at Q4. It does not run the 27B model. Check current price on Amazon.
RTX 4060 Ti 16GB
Runs:
Does not fit:
16 GB does not unlock gemma3:27b. Check current price on Amazon.
RTX 4070 12GB
Runs:
Does not fit:
12 GB matches the Arc B580 for Gemma 3 capacity: gemma3:12b at Q4 (9.2 GB) fits, gemma3:27b does not. The extra 4 GB on a 16 GB card still does not reach 18.2 GB. Check current price on Amazon.
RTX 4070 Ti Super 16GB
Runs:
Does not fit:
Faster than a 4060 Ti, same 16 GB ceiling. gemma3:27b still does not fit. No new price was recorded for this SKU on 11 October.
RTX 4080 16GB
Runs:
Does not fit:
Bandwidth does not change the 18.2 GB miss. Check current price on Amazon.
RTX 3090 24GB (used)
Runs:
Does not fit:
24 GB is the first discrete tier that fits gemma3:27b at Q4. Check current price on Amazon.
AMD RX 7900 XTX 24GB
Runs:
Does not fit:
Same 24 GB ceiling as the 4090 for this model. Not CUDA. Check current price on Amazon.
RTX 4090 24GB
Runs:
Does not fit:
24 GB fits gemma3:27b at Q4 with some KV left. It does not fit Q8. Check current price on Amazon.
RTX 5090 32GB
Runs:
Does not fit:
32 GB fits the 27B Q4 estimate and does not fit the 34.4 GB Q8 estimate. Check current price on Amazon.
Mac mini M6 16GB
Runs:
Does not fit:
Treat usable memory as about 12 GB. gemma3:12b is the ceiling and is tight. gemma3:27b does not fit. Apple lists this 16GB M6 at 153 GB/s (checked 2026-10-18).
Mac mini M6 24GB
Runs:
Does not fit:
A 24 GB Mac is not a 24 GB GPU. About 20 GB is usable after a 4 GB macOS reserve. gemma3:27b at Q4 can load and has little room for a long context. Apple lists this 24GB M6 at 170 GB/s (checked 2026-10-18).
Mac Studio M5 Max 48GB
Runs:
Does not fit:
After a 4 GB macOS reserve a 48 GB M5 Max still holds the 34.4 GB Q8 estimate. Apple lists the 40-core GPU M5 Max at 614 GB/s (checked 2026-10-18). This memory step is an Apple configure.
Mac Studio M5 Max 64GB
Runs:
Does not fit:
64 GB unified holds the Q8 estimate. FP16 at 66.8 GB does not, once macOS is reserved. Apple lists this M5 Max step at 614 GB/s (checked 2026-10-18). It is an Apple configure.
Token generation speed is bottlenecked by memory bandwidth. The table lists vendor bandwidth. This page does not publish tokens per second.
| Hardware | Bandwidth |
|---|---|
| RTX 5090 32GB | 1,792 GB/s |
| RTX 4090 24GB | 1,008 GB/s |
| RTX 4070 Ti Super 16GB | 672 GB/s |
| RTX 4080 16GB | 720 GB/s |
| RTX 4060 Ti 16GB | 288 GB/s |
| Intel Arc B580 12GB | 456 GB/s |
| RTX 4060 8GB | 272 GB/s |
| Mac Studio M5 Max 64GB | 614 GB/s |
| Mac Studio M5 Max 48GB | 614 GB/s |
| Mac mini M5 Pro 24GB | 307 GB/s |
| Mac mini M6 24GB | 170 GB/s |
| Mac mini M6 16GB | 153 GB/s |
Apple bandwidth is from the Mac mini and Mac Studio tech specs, checked 2026-10-18. NVIDIA rows keep the bandwidth figures already used on this site. Tokens per second are not listed.
Ollama
ollama run gemma3:27b Easiest option. One command installs and runs the model. Available tags: gemma3:1b, gemma3:4b, gemma3:12b, gemma3:27b. GPU is auto-detected on NVIDIA, AMD (ROCm), and Apple Silicon. Ollama defaults to Q4_K_M.
LM Studio
Search "Gemma 3" in Discover GUI-based model browser and chat interface. Download GGUF quantizations directly within the app. Best for non-technical users. Runs on Windows, Mac, and Linux. Lets you pick Q4, Q8, or other quants from a dropdown.
Hugging Face + llama.cpp
google/gemma-3-27b-it-GGUF Download GGUF files from the official google org or community repos (bartowski, unsloth) on Hugging Face. Run with llama.cpp for maximum control over quantization, context length, and GPU layer offloading.
For step-by-step installation instructions, see the how to run LLMs locally guide. For a comparison of Ollama vs LM Studio, see the Ollama vs LM Studio guide.
All three are leading open-weight model families. Here is how they differ for local hardware use:
| Model family | Largest dense size | VRAM for largest | Min GPU for largest | Thinking mode | Architecture |
|---|---|---|---|---|---|
| Gemma 3 | 27B | 18.2 GB Q4 | RTX 4090 24GB | No | Dense |
| Qwen3 | 32B | ~17.5 GB Q4 | RTX 4090 24GB | Yes (built-in) | Dense + MoE |
| Llama 3.3 | 70B | ~40 GB Q4 | Dual 24GB GPUs | No | Dense |
Choose Gemma 3 if...
Choose Qwen3 if...
Choose Llama 3.3 if...
RTX 4060 8GB
Fits gemma3:4b at Q4 (4.4 GB) and Q8 (6.8 GB). Does not fit gemma3:12b (9.2 GB). Check current price on Amazon.
Intel Arc B580 12GB
Fits gemma3:12b at Q4 (9.2 GB). Does not fit the 27B model. Confirm the Ollama backend before you buy Arc. Check current price on Amazon.
RTX 4060 Ti 16GB
Does not fit gemma3:27b. 16 GB fits gemma3:12b at Q4 (9.2 GB) and misses 12B at Q8 (16.4 GB). The listing was currently unavailable on 11 October 2026.
RTX 3090 24GB
24 GB fits gemma3:27b at Q4 (18.2 GB) and misses Q8 (34.4 GB). Check current price on Amazon.
RTX 4090 24GB
24 GB fits gemma3:27b at Q4 (18.2 GB) and misses Q8 (34.4 GB) and 70B Q4 (44 GB). Check current price on Amazon.
Mac mini M6 24GB
A 24 GB M6 Mac mini (170 GB/s) can load gemma3:27b at Q4 (18.2 GB) with little room after macOS. It is not a Q8 box. Check current price on Amazon.
For a full cross-budget GPU comparison, see the best GPU for LLMs guide.
What LLMs Can I Run?
Enter your GPU — see every model that fits
Qwen3 Hardware Requirements
Compare Gemma 3 vs Qwen3 for your GPU
Best GPU for LLMs — Full Guide
All budget tiers from entry to workstation
How to Run Gemma 3 Locally
Step-by-step Ollama setup for Gemma 3 1B through 27B — vision included
How to Run LLMs Locally
Step-by-step Ollama, LM Studio, llama.cpp setup
Ollama vs LM Studio
Which tool to use for running Gemma 3 locally
LLM Quantization Explained
Q4 vs Q8 vs FP16 — when trade-offs matter
Apple Silicon for LLMs
M6, M5 Pro, M5 Max: which Mac for Gemma 3?
Best LLM for Coding Locally
Gemma 3 12B and Qwen3 are top coding picks — comparison by GPU tier
VRAM Calculator
Calculate exact VRAM at your context length
Gemma 3 27B at Q4 is 18.2 GB. Buy a 24 GB GPU. A 16 GB card does not fit it. A 24 GB Mac is tight. Q8 is 34.4 GB and does not fit 32 GB. Run ollama run gemma3:27b only when the calculator says the weights fit.
No. 12 × 0.6 + 2 = 9.2 GB, so an 8 GB RTX 4060 does not fit gemma3:12b. Use a 12 GB card, or stay on gemma3:4b. Q8 of the 12B model is 16.4 GB and misses 16 GB.
Both are strong 2025 model families. Gemma 3 27B is one of the best 27B models for reasoning and multilingual tasks, and the Q4 formula is 18.2 GB, so it is a 24 GB model, not a 16 GB model. Qwen3 14B at Q4 is 10.4 GB and does fit 16 GB. Qwen3 32B at Q4 is 21.2 GB and needs 24 GB.
Run: ollama run gemma3:27b — replace "27b" with your target size (1b, 4b, 12b, or 27b). Ollama auto-detects NVIDIA, AMD, and Apple Silicon GPUs and downloads Q4_K_M by default. For an 8 GB GPU, use gemma3:4b. gemma3:12b is 9.2 GB at Q4 and does not fit.
Yes. All Gemma 3 models run on Apple Silicon Macs via Ollama or LM Studio. Unified memory means all RAM is available. A 16 GB Mac can run gemma3:4b and is tight on gemma3:12b. A 24 GB Mac is tight on gemma3:27b at Q4. A 48 GB Mac fits the 27B Q8 estimate of 34.4 GB.
Q4_K_M is the recommended default — it halves VRAM usage versus FP16 with minimal quality loss. Q8_0 offers near-FP16 quality but doubles VRAM requirements. For the 27B model: Q4 is 18.2 GB and Q8 is 34.4 GB. Use Q4 unless you have substantial VRAM headroom and want maximum output quality.
Check VRAM requirements for Gemma 3 models, or compare hardware options.
Apple Silicon for Local LLMs
Mac buyer's guide for running Gemma and other models on M-series chips.
LLM Quantization Explained: Q4, Q8, F16
How quantization lets you fit Gemma 3 into less VRAM.
LM Studio Hardware Requirements
What GPU and RAM you actually need for LM Studio in 2026.
Beginner's Guide to Running AI Locally
Start running open-source models like Gemma 3 in 10 minutes.
VRAM figures on this page use the site formula. Tokens per second are not published here. The formula and the full source list are on the methodology page. For this guide we used:
Spot a number that does not match the linked source? Contact us and we will update the guide.