Gemma 4 12B — Hardware Requirements & GPU Compatibility
Chat
Gemma 4 12B Unified is a base checkpoint in Google DeepMind's Gemma 4 family, an open-weight multimodal model with roughly 12 billion parameters that takes text, image, and audio input and produces text output. Its "Unified" design skips separate encoders per modality, projecting raw image patches and audio directly into the language model, cutting multimodal latency. As a pretrained release, it's a base for fine-tuning or research rather than direct chat use. It supports a 256K token context window and multilingual pretraining across 140+ languages, and is released under the Apache 2.0 license. At roughly 12 billion parameters, it runs on a single consumer GPU with around 8-12 GB of VRAM once quantized to 4-bit.
Runs great on a 24 GB GPU
Runs from 8 GB (Q3_K_L · 16K)Best on 24 GB (Q4_K_M · ~80 tok/s · 128K)Macs from 16 GB (Q4_K_M · ~17 tok/s)
Apache 2.0
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Gemma 4 12B
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run hf.co/DhruvalLabs/gemma-4-12B-GGUF:Q4_K_MQ4_K_M fits fully on the RTX 4090 · 15.0 GB free · 16K context
Download: gemma-4-12B-Q4_K_M.gguf7.38 GBDhruvalLabs
Ways to runOther ways
each opens step-by-stepRun in cloud
Your own private copy · Q4_K_M · 24 GB GPU
- The easiest way to try this model
- Running in a few minutes, easy setup
- Chat right here or connect any app
$5+ bonus for new users
Runs on RTX 3080 Ti (12 GB) (3 GB headroom) · Q4_K_M
- Generation speed
- ~72 tok/s
- generation speed
- Cost per 1M output tokens
- $0.47
- per 1M output tokens
How Much VRAM Does Gemma 4 12B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 6.1 GB | 8.2 GB | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 128K context · ~107 tok/s | |
| 6.3 GB | 8.4 GB | 3-bit small — 3-bit small quantization | Fits: 128K context · ~104 tok/s | |
| 6.9 GB | 9 GB | 3-bit medium — 3-bit medium quantization | Fits: 128K context · ~95 tok/s | |
| 7.2 GB | 9.2 GB | 3-bit large — 3-bit large quantization | Fits: 128K context · ~91 tok/s | |
| 8.2 GB | 10.3 GB | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 128K context · ~80 tok/s | |
| 9.6 GB | 11.6 GB | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 128K context · ~68 tok/s | |
| 10.9 GB | 13 GB | 6-bit, very good — 6-bit quantization, very good quality | Fits: 128K context · ~60 tok/s | |
| 13.0 GB | 15.1 GB | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 128K context · ~50 tok/s |
Selected Q4_K_M
4-bit medium quantization9.0GBat 16K context
- Weights
- 7.2 GB
- KV cache
- 1.6 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 15.0 GB free · ~80 tok/s
≥50 tok/s needs ≥632 GB/s · RTX 5090 ~142, RTX 3090 Ti ~80
Which Hardware Can Run Gemma 4 12B?
Q4_K_M · 8.2 GBGemma 4 12B (Q4_K_M) requires 8.2 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 11+ GB is recommended. Using the full 262K context window can add up to 19.7 GB, bringing total usage to 27.9 GB. 40 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 3080 Ti. 150 devices with unified memory can run Gemma 4 12B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB).
12 GB
12 GB tier →fits · 3.0 GB free · up to 32K ctx
10 GPUs
RTX 3080 Ti~72Simulate output on RTX 3080 TiRTX 5070~53Simulate output on RTX 5070RTX 4070~40Simulate output on RTX 407020 GB
fits · 11.0 GB free · up to 128K ctx
1 GPU
RX 7900 XT~58Simulate output on RX 7900 XT32–40 GB
fits · 23.0 GB free · up to 256K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~142Simulate output on RTX 5090M5 Max36 GB~39Simulate output on Mac Studio M5 Max (36 GB)M532 GB~13Simulate output on MacBook Air 13" M5 (32 GB)Orin~16Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~11Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Alternatives to Gemma 4 12B
Gemma 4 family · 35 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Mistral Nemo Base 2407Mistral AI | — | 11 GB | ~81 | same size |
| Falcon 11BTII UAE | — | 8.5 GB | ~89 | same size |
| Bielik 11B V3.0 Instructspeakleash | — | 27.4 GB | —needs > 24 GB | same size |
| NVIDIA Nemotron Nano 12B v2NVIDIA | — | 11 GB | ~81 | same size |
| Midm 2.0 Base InstructK-intelligence | — | 10.5 GB | ~86 | same size |
| MN 12B Mag Mell R1inflatebot | — | 11 GB | ~81 | same size |
Fine-tunes & variants(3)
Specifications
- Publisher
- Family
- Gemma 4
- Parameters
- 12.0B
- Architecture
- Gemma4UnifiedForConditionalGeneration
- Context Length
- 262,144 tokens
- Vocabulary Size
- 262,144
- Release Date
- 2026-05-23
- License
- Apache 2.0
Frequently Asked Questions
How much VRAM does Gemma 4 12B need?
Gemma 4 12B requires 8.2 GB of VRAM at Q4_K_M, or 25.0 GB at BF16. Full 262K context adds up to 19.7 GB (27.9 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 12.0B × 4.8 bits ÷ 8 = 7.2 GB
KV Cache + Overhead ≈ 1 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 20.7 GB (at full 262K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Gemma 4 12B?
Yes, at Q8_0 (13.0 GB) or lower. Higher quantizations like BF16 (25.0 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Gemma 4 12B?
For Gemma 4 12B, Q4_K_M (8.2 GB) offers the best balance of quality and VRAM usage. Q5_K_S (9.3 GB) provides better quality if you have the VRAM. The smallest option is Q2_K at 6.1 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Gemma 4 12B on a Mac?
Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run Gemma 4 12B. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 8.2 GB of usable unified memory (RAM minus macOS overhead).
Can I run Gemma 4 12B locally?
Yes — Gemma 4 12B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Gemma 4 12B?
At Q4_K_M, Gemma 4 12B can reach ~583 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~80 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 8.23 × 0.65 = ~632 tok/s
Estimated speed at Q4_K_M (8.2 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Gemma 4 12B?
At Q4_K_M, the download is about 7.18 GB. The full-precision BF16 version is 23.92 GB. The smallest option (Q2_K) is 5.08 GB.
Which GPUs can run Gemma 4 12B?
40 consumer GPUs can run Gemma 4 12B at Q4_K_M. Top options include AMD Radeon RX 6800, AMD Radeon RX 6800 XT, AMD Radeon RX 6900 XT, AMD Radeon RX 6700 XT. 26 GPUs have plenty of headroom for comfortable inference.
Which devices can run Gemma 4 12B?
150 devices with unified memory can run Gemma 4 12B at Q4_K_M (8.2 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.