Research Report v2  ·  April 5, 2026  ·  3-Way Multimodal Benchmark

Gemma 4 E2B vs Phi-4 vs Qwen3.5-4B

Text · Vision · Audio · Power · Thermal

A cross-platform deployment comparison of three open-weights multimodal LLMs across two distinct hardware platforms — NVIDIA RTX 3090 (Rocky Linux) and NVIDIA DGX Spark running NVIDIA DGX OS. Measures throughput, latency, VRAM, temperature, and power consumption.

RTX 3090 24GB · Rocky Linux DGX Spark · NVIDIA DGX OS 9 test cases (text + vision + audio) GPU thermal polling every 5s
Gemma 4 E2B — 41.1 tok/s avg · 2.78s latency
Phi-4 Multimodal — 33.9 tok/s avg · 3.76s latency
Qwen3.5-4B — 15.2 tok/s avg · 8.73s latency · 26W power
1

Model Overview

Gemma 4 E2B
Google DeepMind
HF IDgoogle/gemma-4-E2B-it
Effective params2.3B active
Total params5.1B (w/ embeddings)
ArchitectureDense + PLE
Context window128K tokens
ModalitiesText · Image · Audio · Video
VRAM (BF16)10.3 GB
Thinking mode✓ Native
Deployment roleThroughput frontier
Distinctive edgeOnly video-capable model here
Phi-4 Multimodal
Microsoft · February 2025
HF IDmicrosoft/Phi-4-multimodal-instruct
Parameters~5.6B
ArchitectureDense Transformer
Context window128K tokens
ModalitiesText · Image · Audio
VRAM (FP16)11.6 GB
Thinking modeNot reported
OpenASR WER#1 · 6.14%
Deployment roleASR / concise-output niche
Qwen3.5-4B
Qwen Team
HF IDQwen/Qwen3.5-4B
Parameters4B
ArchitectureGated DeltaNet + MoE
Context window262K tokens
ModalitiesText · Image
HardwareGB10 DGX Spark
Thinking mode✓ Native
Deployment roleEfficiency frontier
Distinctive edgeLargest context window
Hardware split: Gemma 4 E2B and Phi-4 Multimodal run sequentially on Rocky Linux (NVIDIA RTX 3090 24 GB, CUDA 12.4, Linux). Qwen3.5-4B runs separately on NVIDIA DGX OS (NVIDIA DGX Spark, ARM aarch64, CUDA 13, BF16 via Docker).
2

Hardware Platform Comparison

NVIDIA RTX 3090 (Rocky Linux)

ArchitectureAmpere GA102
CUDA Cores10,496
VRAM24 GB GDDR6X
Memory Bandwidth936 GB/s
TDP (rated)350 W
Power draw (idle)~17 W
Power draw (inference)157–205 W avg
Peak power (this run)239.8 W
Temp peak (Phi-4)58 °C
Temp peak (Gemma4)61 °C

NVIDIA DGX Spark (NVIDIA DGX OS)

ArchitectureBlackwell GB10 SoC
CPU20-core ARM Grace
Unified Memory128 GB LPDDR5X
GPU Memory BW273 GB/s
L2 / L4 Cache24 MB / 16 MB
System TDP~140 W total
GPU power draw (idle)~11 W
GPU power draw (inference)25.8 W avg
Peak power (this run)27.1 W
Temp peak (Qwen)51 °C
Key contrast: The RTX 3090 is a traditional discrete GPU with dedicated GDDR6X VRAM and high thermal budget (TDP 350W). The GB10 is a novel Blackwell SoC integrating CPU+GPU+memory in a single package with dramatically lower power draw (~26W GPU portion vs 157–205W on RTX 3090). The trade-off is lower memory bandwidth (273 vs 936 GB/s), which directly affects token generation speed.
3

Generation Speed (tokens/second)

■ Gemma 4 E2B ■ Phi-4 ■ Qwen3.5-4B max 52 tok/s shown
Text
Basic Q&A
17.6
32.9
8.3
Reasoning
49.4
38.1
17.5
Coding
49.5
38.1
17.3
Multilingual
50.0
38.2
12.0
Summarization
49.3
37.8
16.6
Vision
Image Description
48.7
28.2
17.5
Color Identification
47.2
24.6
17.2
Audio (Phi-4 & Gemma4 only — Qwen3.5 text+image only)
Transcription
13.9
33.4
Audio Q&A
44.0
33.4
4

Power Draw & Thermal Profile

Average Power During Inference

Scale: 0 → 240 W
Gemma4 / RTX 3090
204.8 W avg
Phi-4 / RTX 3090
156.5 W avg
Qwen / GB10
25.8 W avg

The GB10 (DGX Spark) draws 8× less power than Gemma4 on RTX 3090 during inference. For always-on or edge AI deployments, this translates to dramatically lower electricity costs and no need for active cooling.

Peak Temperature During Inference

Scale: 0 → 70 °C
Gemma4 / RTX 3090
61.0 °C peak
Phi-4 / RTX 3090
58.0 °C peak
Qwen / GB10
51.0 °C peak

All systems remain well within safe operating limits. The GB10 runs 10°C cooler despite its integrated SoC design. RTX 3090 temperatures are comfortable at 58–61°C, leaving significant thermal headroom before throttling (~83°C typical limit).

Model Hardware Avg tok/s Avg Power (W) Peak Power (W) tok/s per Watt Temp (avg / peak) VRAM Used GPU Util %
Gemma 4 E2B RTX 3090 41.1 204.8 W 227.3 W 0.20 58.8°C / 61°C 10.3 GB 57%
Phi-4 Multimodal RTX 3090 33.9 156.5 W 239.8 W 0.22 55.4°C / 58°C 11.6 GB 30%
Qwen3.5-4B GB10/DGX 15.2 25.8 W 27.1 W 0.59 47.7°C / 51°C N/A* 90%
* DGX Spark uses unified memory — nvidia-smi does not report GPU-only VRAM. Qwen3.5-4B requires ~8 GB LPDDR5X of the 128 GB pool.

tok/s per Watt (power efficiency): Qwen on GB10 is 3× more power-efficient than Gemma4 on RTX 3090 (0.59 vs 0.20 tok/s/W). However, raw throughput is 2.7× lower — so the optimal choice depends on latency vs cost requirements.
5

Head-to-Head Results

Modality Test Phi-4 tok/s Gemma 4 tok/s Qwen3.5 tok/s Phi-4 lat. Gemma4 lat. Qwen lat. Accuracy Winner (speed)
Text Basic Q&A 32.9 17.6 8.3 2.24s 0.68s 1.20s All ✓ Gemma4 (latency)
Text Reasoning 38.1 49.4 17.5 8.02s 6.57s 17.15s All correct† Gemma4
Text Python coding 38.1 49.5 17.3 3.94s 5.22s 14.45s All ✓ Gemma4
Text Multilingual 38.2 50.0 12.0 1.77s 3.05s 7.41s All ✓ Gemma4
Text Summarization 37.8 49.3 16.6 2.75s 1.33s 3.61s All ✓ Gemma4
Vision Image Description 28.2 48.7 17.5 2.88s 4.22s 11.45s All ✓ Gemma4
Vision Color ID 24.6 47.2 17.2 1.13s 2.13s 5.83s All ✓ Gemma4
Audio Transcription 33.4 13.9 N/A 6.13s 0.44s Both ✓ Phi-4 (tok/s)
Audio Audio Q&A 33.4 44.0 N/A 5.02s 1.38s Both ✓ Gemma4
Reasoning (snail puzzle): All three models produce correct reasoning chains but are truncated at the benchmark token limit before printing the final short answer. This is a benchmark artifact, not a reasoning failure.
6

VRAM / Memory Efficiency

VRAM at Rest (model loaded, RTX 3090)

Gemma 4 E2B
10.3 GB
Phi-4
11.6 GB
Qwen3.5-4B *
~8 GB unified

Gemma 4 E2B uses 1.3 GB less VRAM than Phi-4 despite a larger total parameter count. The Per-Layer Embedding (PLE) architecture keeps embedding tables in a separate lookup space, reducing active computation memory. Qwen3.5-4B on GB10 draws from a shared 128 GB unified pool.

Peak VRAM During Inference

Gemma4: 10.8 GB peak
13.8 GB free
Phi-4: 11.9 GB peak
12.7 GB free
8 GB
Qwen: ~8 GB / 128 GB unified (6% usage)

Qwen3.5-4B uses only ~6% of the GB10's 128 GB unified memory pool — leaving massive headroom for very long context (262K tokens) or multi-model serving.

7

Response Quality Comparison

Text — Summarization (AI definition passage)
Gemma 4 E2B
"Artificial intelligence (AI) is the demonstration of intelligence in machines, contrasting with natural intelligence found in animals and humans. The field of AI research focuses on intelligent agents — systems that perceive their environment and act to maximize their chances of achieving specific goals."
Phi-4 Multimodal
"Artificial intelligence (AI) refers to the intelligence demonstrated by machines, distinct from natural intelligence, and involves the study of intelligent agents that perceive their environment and act to achieve their goals. The term 'artificial intelligence' also encompasses machines that simulate human cognitive abilities like learning and problem-solving."
Qwen3.5-4B
"Artificial intelligence is the field of study focused on creating intelligent agents that perceive their environments and take actions to achieve specific goals. While the term originally described machines mimicking human cognitive skills like learning and problem-solving, it now broadly encompasses any machine intelligence distinct from natural human or animal intelligence."
All three produce high-quality, accurate 2-sentence summaries. Gemma4 is most concise, Phi-4 is plain and direct, and Qwen is more expansive while remaining accurate.
Vision — Image Description (synthetic house scene)
Gemma 4 E2B
"This is a simple, stylized illustration with a distinct, childlike aesthetic.

Sky: Clear light blue, upper portion.
Sun: Bright solid yellow circle, upper right.
Ground: Solid green block (grass/field).
House: Centered. Reddish-orange triangular roof. Light beige walls. Two square light-blue windows. Brown door."
Phi-4 Multimodal
"The image depicts a simple, stylized representation of a house. The house has a red roof and a brown door, with two windows on either side of the door. The house is situated on a green grassy field, and there is a yellow sun in the sky above the house. The background is a light blue sky."
Qwen3.5-4B
"This is a simple, stylized digital illustration of a small house in a rural or suburban setting. The image is composed of flat, solid colors and basic geometric shapes, giving it a minimalist or childlike aesthetic.

Sky: Light blue. Sun: Solid yellow disc, top-right. Ground: Solid green. House: Red triangular roof. Beige walls. Two windows. Brown door."
Audio — Synthetic Tone Test (3s 200–800Hz sine wave)

The synthetic audio file (3s, 200–800Hz mixed tones) is not real speech — it generates ambiguous transcription results from all models. Phi-4 loops "Oh. [no audio]..." (200 tokens), Gemma4 correctly identifies minimal content ("Oh,"). Both handle real speech well per the v1 benchmark with KOKORO TTS. Qwen3.5-4B does not have audio capability in the current deployment.

Audio verdict: Gemma4 wins on latency (0.44s vs 6.1s for transcription). For real speech, both models perform comparably. Phi-4 loops on ambiguous input; Gemma4 terminates cleanly.

8

Key Findings

Gemma 4 E2B is fastest: 21% over Phi-4, 170% over Qwen on sustained text

For text tasks requiring more than 50 tokens, Gemma 4 averages ~49.5 tok/s vs Phi-4's ~38.1 tok/s (+30%) and Qwen3.5-4B's ~15.2 tok/s (+226%). The RTX 3090's 936 GB/s memory bandwidth and Gemma4's efficient PLE architecture combine for highest throughput. Gemma4 also wins on latency: average 2.78s vs 3.76s (Phi-4) and 8.73s (Qwen).

🌱

Qwen3.5-4B on GB10 defines the efficiency frontier: 0.59 tok/s/W

Despite lower throughput, Qwen on the DGX Spark uses only 25.8W average (vs 156–205W on RTX 3090). Its token-per-watt ratio (0.59) is nearly 3× better than Gemma4 (0.20) and Phi-4 (0.22). In this benchmark, DGX Spark plus Qwen establishes a strong low-power operating point for edge-style or always-on deployment. GPU utilization is consistently 90%, showing the platform is heavily used at a fraction of the energy cost of the RTX 3090 runs.

👁️

Vision: Gemma4 is 1.7× faster than Phi-4, 2.8× faster than Qwen

Gemma4 vision throughput (~48 tok/s) dwarfs Phi-4 (~26 tok/s avg) and Qwen (~17 tok/s). All three produce accurate, detailed scene descriptions. Gemma4 uses structured markdown with bold headers; Phi-4 gives concise prose; Qwen produces thorough descriptions with clear section breaks. Quality is comparable across models — Gemma4 wins purely on speed.

🧊

GB10 runs 10°C cooler — zero active cooling needed at 27W peak

The DGX Spark's GB10 SoC peaked at 51°C during 7 consecutive inference calls. The RTX 3090 reached 58–61°C for the same workload. At 27W peak power draw, the GB10 can run fanless-equivalent workloads. This opens up deployment scenarios in noise-sensitive environments (clinical, office, embedded) where the RTX 3090's fan noise and thermal management would be problematic. The 140W total DGX Spark system TDP easily fits on a desk-sized workstation.

📐

Context windows: Qwen 262K > Gemma4 128K = Phi-4 128K

Qwen3.5-4B's 262K context window is the largest of the three, suitable for entire codebases, legal documents, or very long conversations. Gemma4 and Phi-4 both expose 128K-token context windows in current official documentation. That still leaves Qwen in the lead for long-document workloads, but Phi-4 should not be treated as a short-context outlier.

🎙️

Audio modality: Gemma4 and Phi-4 only — Qwen3.5 is text+image

Qwen3.5-4B as deployed on NVIDIA DGX OS does not support audio input. Gemma4 and Phi-4 both handle audio, with Gemma4 having dramatically lower end-to-end latency (0.44s vs 6.1s for transcription) and winning audio Q&A throughput (44.0 vs 33.4 tok/s). For full multimodal pipelines requiring audio, Phi-4 or Gemma4 on the RTX 3090 are the options.

🏗️

Architecture diversity: three different AI chip designs, three different trade-offs

This benchmark spans three distinct compute paradigms: (1) RTX 3090 — traditional discrete GPU, high bandwidth, high power, maximum throughput; (2) DGX Spark — Blackwell SoC, unified memory, ultra-low power, edge AI; (3) Three model architectures: Dense+PLE (Gemma4), Dense Transformer (Phi-4), Gated DeltaNet+MoE (Qwen3.5). The 2026 AI ecosystem is no longer "one architecture wins" — deployment target determines the optimal choice.

🎬

Video understanding remains Gemma 4 exclusive — neither Phi-4 nor Qwen3.5 can do it

Gemma 4 E2B is the only model in this comparison with reported video understanding capability. In the benchmark, an 8-second 1280×720 MP4 sampled at 1 fps was analyzed in 13.3s at 42.9 tok/s. Phi-4 has no video support in this benchmark. Qwen3.5-4B as deployed has no video capability. For video AI workflows, Gemma 4 E2B on the RTX 3090 is the only option in this comparison.

9

Power Efficiency Deep Dive

Qwen / GB10
0.59
tok/s per Watt
Best power efficiency
Phi-4 / RTX 3090
0.22
tok/s per Watt
Moderate efficiency
Gemma4 / RTX 3090
0.20
tok/s per Watt
Highest throughput

Estimated Monthly Energy Cost (24/7 operation)

Setup Avg Power kWh / month Cost @ $0.15/kWh Tokens @ 15.2 avg tok/s
Qwen3.5-4B / GB10 25.8 W 18.6 kWh $2.79 / month 39.3 B tokens
Phi-4 / RTX 3090 156.5 W 112.7 kWh $16.91 / month 87.4 B tokens
Gemma4 / RTX 3090 204.8 W 147.5 kWh $22.12 / month 106.1 B tokens

Assumes full utilization 24/7. Real workloads are bursty — idle power is much lower (17W for RTX 3090, 11W for GB10). For inference-heavy production workloads, the DGX Spark's 8× lower power draw compounds into significant operational savings, especially when scaled across multiple units.

10

Verdict & Recommendations

No Single Global Winner

The corrected picture is a split deployment frontier. Gemma 4 E2B defines the throughput frontier, Qwen3.5-4B on DGX Spark defines the efficiency frontier, and Phi-4 remains relevant for concise text-image-audio pipelines and strong ASR throughput.

Use Gemma 4 E2B when…

  • High-throughput text generation needed (49+ tok/s)
  • Vision tasks are core — nearly 2× faster than Phi-4, 2.8× faster than Qwen
  • Video analysis required — exclusive capability
  • Thinking/reasoning mode improves answer quality
  • Long context (128K) plus full modality coverage needed
  • You have an RTX 3090/4090 and want max performance

Use Phi-4 Multimodal when…

  • Audio transcription throughput matters
  • Short response latency is less critical
  • Pipeline is already built on Phi-4 prompt format
  • Concise prose over markdown is preferred
  • Text, image, and audio support on one RTX workflow is enough

Use Qwen3.5-4B / GB10 when…

  • Power budget is critical (25W vs 200W)
  • Always-on 24/7 deployment at minimal electricity cost
  • Extremely long context needed (262K tokens)
  • ARM/edge deployment or noise-sensitive environments
  • Throughput of 15 tok/s is sufficient for your use case
  • Multi-model serving (128 GB unified memory)


Next steps: (1) Run same-model cross-hardware replications to separate model effects from platform effects. (2) Add Qwen audio via a dedicated pipeline to enable a true 3-way audio comparison. (3) Benchmark different context lengths to find where DGX Spark's memory profile becomes more attractive than RTX throughput. (4) Test batch inference, where 128 GB unified memory may help. (5) Add first-token latency and larger standardized multimodal evaluation sets.
11

Infrastructure & Reproducibility

Server: Rocky Linux (Phi-4 + Gemma4)

GPUNVIDIA RTX 3090 24 GB
RAM32 GB DDR4
CUDA12.4
Phi-4 dtypeauto (FP16/BF16)
Gemma4 dtypeBF16
Port9000 (both, sequential)
FrameworkFlask + Transformers

Server: NVIDIA DGX OS (Qwen3.5-4B)

GPUNVIDIA GB10 (DGX Spark)
CPUARM Grace 20-core
Memory128 GB LPDDR5X unified
CUDA13 (Blackwell)
Qwen dtypeBF16
Port8002
FrameworkFastAPI + uvicorn + Docker