Skip to content
Updated minutes ago
Google

PaLI-Gemma 3B

Google · dense · 2.9B parameters · 8,192 context

Quality
50.0

Parameters

2.9B

Context Window

8K tokens

Architecture

Dense

Best GPU

RTX 4070 Ti

Intelligence Brief

PaLI-Gemma 3B is a 2.9B parameter DENSE model from Google, featuring Multi-Head Attention (MHA) with 18 layers and 2,048 hidden dimensions. With a 8,192 token context window, it supports vision, multilingual. For self-hosted inference, RTX 4070 Ti delivers optimal throughput at $237/month.

Architecture Details

TypeDENSE
Total Parameters2.9B
Active Parameters2.9B
Layers18
Hidden Dimension2,048
Attention Heads8
KV Heads8
Head Dimension256
Vocab Size257,152

Memory Requirements

BF16 Weights

5.8 GB

FP8 Weights

2.9 GB

INT4 Weights

1.4 GB

KV-Cache per Token73728 bytes
Activation Estimate0.40 GB

GPU Compatibility Matrix

PaLI-Gemma 3B is compatible with 100% of GPU configurations across 41 GPUs at 3 precision levels.

BF16 (Full)
FP8 (Half)
INT4 (Quarter)
Blackwell(7 GPUs)
B200 NVL (pair)360GB
B300288GB
B100 SXM192GB
GB200 NVL72 (per GPU)192GB
Hopper(7 GPUs)
H100 NVL 94GB (per GPU pair)188GB
H200 SXM141GB
H2096GB
GH20096GB
Ada Lovelace(11 GPUs)
L40S48GB
L4048GB
RTX 6000 Ada48GB
L2048GB
Ampere(16 GPUs)
A100 80GB SXM80GB
A100 80GB PCIe80GB
A1664GB
RTX A600048GB
Legend:No fitVery tightTightModerateGoodExcellent

GPU Recommendations

RTX 4070 Tioptimal

BF16 · 1 GPU · vllm

100/100

score

Throughput

422.3 tok/s

Latency (ITL)

2.4ms

Est. TTFT

0ms

Cost/Month

$237

Cost/M Tokens

$0.21

Use this config →
RTX 3080optimal

BF16 · 1 GPU · vllm

100/100

score

Throughput

636.8 tok/s

Latency (ITL)

1.6ms

Est. TTFT

0ms

Cost/Month

$133

Cost/M Tokens

$0.08

Use this config →
RTX 4060optimal

BF16 · 1 GPU · vllm

100/100

score

Throughput

227.9 tok/s

Latency (ITL)

4.4ms

Est. TTFT

1ms

Cost/Month

$209

Cost/M Tokens

$0.35

Use this config →

Deployment Options

API

API Deployment

No API pricing available

Self-Hosted

Single GPU

RTX 4070 Ti

$237/mo

Min VRAM: 3 GB

Scale

Multi-GPU

RTX 4070 Ti

422.3 tok/s

Best available config

API Pricing Comparison

No API pricing data available for this model.

Performance Estimates

Throughput by GPU

RTX 4070 Ti
422.3 tok/s
RTX 3080
636.8 tok/s
RTX 4060
227.9 tok/s

VRAM Breakdown (RTX 4070 Ti, BF16)

Weights
KV
Act
Weights 5.8 GBKV-Cache 2.4 GBActivations 3.2 GBOverhead 0.5 GB

Precision Impact

bf16

5.8 GB

weights/GPU

~422.3 tok/s

fp8

2.9 GB

weights/GPU

int4

1.4 GB

weights/GPU

Capabilities

Features

Tool Use Vision Code Math Reasoning Multilingual Structured Output

Supported Frameworks

vllmtgi

Supported Precisions

BF16 (default)FP8INT4

Where to Deploy PaLI-Gemma 3B

Similar Models

Frequently Asked Questions

How much VRAM does PaLI-Gemma 3B need for inference?

PaLI-Gemma 3B requires approximately 5.8 GB of VRAM at BF16 precision, 2.9 GB at FP8, or 1.4 GB at INT4 quantization. Additional VRAM is needed for KV-cache (73728 bytes per token) and activations (~0.40 GB).

What is the best GPU for PaLI-Gemma 3B?

The top recommended GPU for PaLI-Gemma 3B is the RTX 4070 Ti using BF16 precision. It achieves approximately 422.3 tokens/sec at an estimated cost of $237/month ($0.21/M tokens). Score: 100/100.

How much does PaLI-Gemma 3B inference cost?

PaLI-Gemma 3B inference costs vary by provider and GPU setup. Use our calculator for detailed cost estimates across all providers.