Skip to content
Updated minutes ago
Microsoft

Phi 3.5 Vision

Microsoft · dense · 4.2B parameters · 131,072 context

Quality
50.0

Parameters

4.2B

Context Window

128K tokens

Architecture

Dense

Best GPU

A4000

Intelligence Brief

Phi 3.5 Vision is a 4.2B parameter DENSE model from Microsoft, featuring Multi-Head Attention (MHA) with 32 layers and 3,072 hidden dimensions. With a 131,072 token context window, it supports vision, structured output, code, math. For self-hosted inference, A4000 delivers optimal throughput at $161/month.

Architecture Details

TypeDENSE
Total Parameters4.2B
Active Parameters4.2B
Layers32
Hidden Dimension3,072
Attention Heads32
KV Heads32
Head Dimension96
Vocab Size32,064

Memory Requirements

BF16 Weights

8.4 GB

FP8 Weights

4.2 GB

INT4 Weights

2.1 GB

KV-Cache per Token393216 bytes
Activation Estimate0.50 GB

GPU Compatibility Matrix

Phi 3.5 Vision is compatible with 98% 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

A4000optimal

BF16 · 1 GPU · vllm

100/100

score

Throughput

259.2 tok/s

Latency (ITL)

3.9ms

Est. TTFT

1ms

Cost/Month

$161

Cost/M Tokens

$0.24

Use this config →
RTX 4080optimal

BF16 · 1 GPU · vllm

100/100

score

Throughput

414.8 tok/s

Latency (ITL)

2.4ms

Est. TTFT

0ms

Cost/Month

$304

Cost/M Tokens

$0.28

Use this config →
RTX 4070 Tioptimal

BF16 · 1 GPU · vllm

100/100

score

Throughput

291.6 tok/s

Latency (ITL)

3.4ms

Est. TTFT

1ms

Cost/Month

$237

Cost/M Tokens

$0.31

Use this config →

Deployment Options

API

API Deployment

No API pricing available

Self-Hosted

Single GPU

A4000

$161/mo

Min VRAM: 4 GB

Scale

Multi-GPU

RTX 3070 x2

375.3 tok/s

TP· $171/mo

API Pricing Comparison

No API pricing data available for this model.

Performance Estimates

Throughput by GPU

A4000
259.2 tok/s
RTX 4080
414.8 tok/s
RTX 4070 Ti
291.6 tok/s

VRAM Breakdown (A4000, BF16)

Weights
KV
Act
Weights 8.4 GBKV-Cache 6.4 GBActivations 4.0 GBOverhead 0.7 GB

Precision Impact

bf16

8.4 GB

weights/GPU

~259.2 tok/s

fp8

4.2 GB

weights/GPU

int4

2.1 GB

weights/GPU

Capabilities

Features

Tool Use Vision Code Math Reasoning Multilingual Structured Output

Supported Frameworks

vllmsglangtgitensorrt-llmollama

Supported Precisions

BF16 (default)FP8INT4

Where to Deploy Phi 3.5 Vision

Similar Models

Gemma 3 4B

4.3B params · dense

Quality: 54

from $0.10/M

Similar specsCompare →

Minitron 4B

4B params · dense

Quality: 50

from $0.06/M

Smaller contextCompare →

Nemotron Mini 4B

4B params · dense

Quality: 48

from $0.06/M

Smaller contextCompare →

Qwen 3 4B

4B params · dense

Quality: 57

from $0.10/M

Higher qualityCompare →

Frequently Asked Questions

How much VRAM does Phi 3.5 Vision need for inference?

Phi 3.5 Vision requires approximately 8.4 GB of VRAM at BF16 precision, 4.2 GB at FP8, or 2.1 GB at INT4 quantization. Additional VRAM is needed for KV-cache (393216 bytes per token) and activations (~0.50 GB).

What is the best GPU for Phi 3.5 Vision?

The top recommended GPU for Phi 3.5 Vision is the A4000 using BF16 precision. It achieves approximately 259.2 tokens/sec at an estimated cost of $161/month ($0.24/M tokens). Score: 100/100.

How much does Phi 3.5 Vision inference cost?

Phi 3.5 Vision inference costs vary by provider and GPU setup. Use our calculator for detailed cost estimates across all providers.