Google

Gemma 4 12B

vision

12B

Parameters

7,1 Go

Q4_K_M file

≥ 8,6 Go

Recommended VRAM

262 144

Max context

Standard llama-bench run (prompt 512 / generation 128 tokens), model fully in VRAM. Click a card for details. Electricity cost at the default price (0,25 €/kWh).

Card · Q4_K_MVRAMGenerationPromptVRAM usedPower€ elec. / M tokPrice
RTX 4070 Ti12 Go53,8 tok/s3 8058,5 Go161 W0,21 €/M tok—

Not measured on these cards yet — click for an estimate:

Technical sheet

Published
2026-05-23
Architecture
48 layers · 8 KV heads · dimension 256
Attention
Sliding window: 8 of 48 layers keep the whole context, the others only the last 1024 tokens — a much lighter cache.
Context cache (f16)
≈ 63 MB per 1,000 tokens · 2,0 Go for 32,768 tokens
GGUF file
bartowski/gemma-4-12B-it-GGUF · gemma-4-12B-it-Q4_K_M.gguf Download (7,1 Go)

Public Hugging Face data (official config.json, GGUF repo). The context cache only counts full-attention layers.