OCR / VisionGemmaMIT

Gemma 3 12B

google/gemma-3-12b-it

Gemma 3 12B, fits single consumer GPU.

Parameters

12B

Context

128K

Downloads

520K

Likes

2.1K

Architecture

transformer

View on HuggingFace

GPU Requirements

VRAM / RAM / disk estimates by quantization and usage scenario

QuantizationMinimumRecommendedProduction
FP1616bit
33.2 GB
RAM 49.8 · Disk 73.9
50.3 GB
RAM 75.4 · Disk 99.9
148.5 GB
RAM 222.7 · Disk 138.9
FP88bit
19.1 GB
RAM 32.0 · Disk 56.4
35.5 GB
RAM 53.2 · Disk 82.4
133.0 GB
RAM 199.5 · Disk 121.4
INT88bit
19.1 GB
RAM 32.0 · Disk 56.4
35.5 GB
RAM 53.2 · Disk 82.4
133.0 GB
RAM 199.5 · Disk 121.4
AWQ4bit
12.3 GB
RAM 32.0 · Disk 48.0
28.4 GB
RAM 42.6 · Disk 74.0
125.6 GB
RAM 188.4 · Disk 113.0
GPTQ4bit
12.4 GB
RAM 32.0 · Disk 48.2
28.5 GB
RAM 42.7 · Disk 74.2
125.7 GB
RAM 188.6 · Disk 113.2
GGUF4bit
12.6 GB
RAM 32.0 · Disk 48.4
28.6 GB
RAM 43.0 · Disk 74.4
125.9 GB
RAM 188.8 · Disk 113.4

GPU Recommendations

Best ValueBEST

1× H100 SXM

80 GB VRAM

$1,170/mo

≈ 577.6 tok/s

Effective production throughput ≈ 404.3 tok/s

score 0.5841

Deploy on vast ↗
Cheapest

2× Tesla V100

32 GB VRAM

$40.6/mo

≈ 0 tok/s (N/A — benchmark unavailable)

Effective production throughput ≈ 0 tok/s

score 0.6102

Deploy on vast ↗
Performance

8× H200 SXM

1128 GB VRAM

$20,440/mo

≈ 2648.3 tok/s

Effective production throughput ≈ 1853.8 tok/s

score 0.2197

Min Complexity

1× RTX A6000

48 GB VRAM

$210/mo

≈ 132.4 tok/s

Effective production throughput ≈ 92.7 tok/s

score 0.6487

Deploy on vast ↗

💬 Community — real deployment experience

0 Articles · 0 Benchmarks

View Community →

GPUs that fit (single card)

RTX A6000

48 GB VRAM

$210

/mo · vast

RTX 5090

32 GB VRAM

$236

/mo · vast

L40S

48 GB VRAM

$341

/mo · vast

RTX PRO 6000 WS

48 GB VRAM

$482

/mo · vast

RTX PRO 6000 S

48 GB VRAM

$633

/mo · vast

AmciHub

AI Model Deployment & Compute Intelligence — analyze AI model requirements, GPU performance, cloud pricing and deployment costs to find the right deployment solution.

© 2026 AmciHub. All rights reserved. Model → Requirement → GPU → Benchmark → Cloud → Cost → Recommendation