Advisor / Models / Gemma 4 12B

Gemma 4 12B: VRAM requirements & speed

Gemma 4 12B (11.96B dense, apache-2.0). Native context 256k. Smallest GGUF: 7.12 GB.

7.12 GBsmallest quant file
25hardware pages that fit it fully
1run with offload
10measured data points

Memory needed by quantization and context

GGUF file size plus f16 KV cache (and fixed overhead of ~1 GB per GPU not shown). Source: unsloth/gemma-4-12b-it-GGUF.

QuantFile+8k ctx+32k ctx+128k ctx+256k ctx
Q4_K_M7.12 GB7.6 GB8.0 GB9.6 GB11.8 GB
Q5_K_M8.41 GB8.9 GB9.3 GB10.9 GB13.0 GB
Q6_K9.79 GB10.3 GB10.7 GB12.3 GB14.4 GB
Q8_013.13 GB13.6 GB14.0 GB15.6 GB17.8 GB

KV cache size by context length

Computed from the model's attention layout (sliding-window layers capped at their window).

Contextf16q8_0q4_0
8k0.47 GB0.25 GB0.13 GB
32k0.87 GB0.46 GB0.25 GB
128k2.48 GB1.32 GB0.70 GB
256k4.63 GB2.46 GB1.30 GB

Which hardware runs Gemma 4 12B? · 16k context

HardwareQuantFitMemoryTok/sSpeed source
RTX 5090 32GBQ8_0fits fully13.7 GB91scaled from measured
RTX 4090 24GBQ8_0fits fully13.7 GB64measured
Mac Studio M5 Ultra 256GBQ8_0fits fully13.7 GB58estimated
Mac Studio M5 Ultra 512GB (ships late Oct 2026)Q8_0fits fully13.7 GB58estimated
RTX 3090 24GBQ8_0fits fully13.7 GB57measured
RTX 5070 12GBQ6_Kfits fully10.4 GB51scaled from measured
2x RTX 3090 (48GB total)Q8_0fits fully13.7 GB49estimated
AMD RX 7900 XTX 24GBQ8_0fits fully13.7 GB45scaled from measured
RTX PRO 6000 96GB (bonus page)Q8_0fits fully13.7 GB44scaled from measured
RTX 4070 12GBQ6_Kfits fully10.4 GB43scaled from measured
RTX 5080 16GBQ8_0fits fully13.7 GB39scaled from measured
Mac Studio M5 Max 128GBQ8_0fits fully13.7 GB36scaled from measured
MacBook Pro 14/16 M5 Max 128GBQ8_0fits fully13.7 GB36scaled from measured
RTX 3060 12GB (32GB RAM)Q6_Kfits fully10.4 GB29scaled from measured
Mac Studio (2025) M3 Ultra 96GBQ8_0fits fully13.7 GB24scaled from measured
RTX 5070 Ti 16GBQ8_0fits fully13.7 GB21scaled from measured
8GB VRAM class (RTX 4060 8GB, 32GB RAM)Q4_K_Mpartial offload (slow)7.7 GB21estimated
16GB VRAM class (RTX 4060 Ti 16GB, 32GB RAM)Q8_0fits fully13.7 GB21measured
MacBook Pro 14/16 (2024) M4 Max 128GBQ8_0fits fully13.7 GB20scaled from measured
Mac mini M6 32GBQ4_K_Mfits fully7.7 GB17scaled from measured
NVIDIA DGX Spark 128GBQ6_Kfits fully10.4 GB17scaled from measured
MacBook Pro 14/16 M5 Pro 48GBQ8_0fits fully13.7 GB16scaled from measured
Mac mini (2024) M4 Pro 64GBQ6_Kfits fully10.4 GB16scaled from measured
AMD Ryzen AI Max+ 395 (Strix Halo) 128GBQ4_K_Mfits fully7.7 GB15scaled from measured
MacBook Air M5 32GBQ4_K_Mfits fully7.7 GB13scaled from measured
Mac mini / iMac M4 16GBQ4_K_Mfits fully7.7 GB11scaled from measured

Discrete GPUs are modelled with 32 GB of system RAM available for offload. Fits (at least with offload) on every listed configuration.

Links

Model card · GGUF files · Open in the advisor · last verified 2026-10-14

FAQ

How much VRAM does Gemma 4 12B need?

The smallest quant (Q4_K_M) is 7.12 GB, so you need roughly 9 GB for an 8k context including KV cache and overhead. At 256k context the f16 KV cache alone is 4.6 GB.

Can I run Gemma 4 12B on a RTX 5090 32GB?

Yes: on RTX 5090 32GB it fits fully as Q8_0 at about 91 tok/s (scaled from measured). 25 of 26 hardware configurations fit it fully at 16k context.

What is the maximum context for Gemma 4 12B?

256k tokens natively. Memory, not the model, is usually the limit: see the KV cache table above.

All models: model index · by memory: 8 GB · 12 GB · 16 GB · 24 GB · 32 GB · 48 GB · 64 GB · 96 GB · 128 GB · 192 GB · 256 GB · 512 GB. Numbers use the same fit math as the advisor; sources and methods.