Gemma4-E2B vs Gemma4-E4B / Unsloth
vs Qwen3.5-4B — 3-Way Evaluation

16 structured test cases · text · image · audio · video · quality scored by Claude Sonnet 4.6

Gemma4-E2B · RTX 3090 · bf16 · dlyog04:9000
Gemma4-E4B · GB10 DGX · Unsloth 4-bit · dgx1:9001
Qwen3.5-4B · GB10 DGX · bf16 · dgx1:8002
Performance — April 14 2026 evaluation run
47.5
tok/s · Gemma4-E2B
4.6s
avg latency
16/16
success
bf16
precision
20.2
tok/s · Gemma4-E4B / Unsloth
11.2s
avg latency
14/15
success
4-bit
precision
16.9
tok/s · Qwen3.5-4B
17.7s
avg latency
10/10
success
bf16
precision
Gemma4-E2B
47.5 tok/s
Gemma4-E4B
20.2 tok/s
Qwen3.5-4B
16.9 tok/s

E2B is 2.4× faster than E4B. Speed gap is hardware-driven: RTX 3090 936 GB/s GDDR6X vs GB10 ~273 GB/s LPDDR5X. LLM inference is memory-bandwidth-bound.

Response Quality — Scored by Claude Sonnet 4.6 (AI Evaluator) · 0–10 scale
Modality Cases Gemma4-E2B Gemma4-E4B Qwen3.5-4B
Text5 7.07.0 8.5 🏆
Image5 7.58.0 8.5 🏆
Audio5 8.5 🏆 7.5 1 failure N/A
Video1 7.0 N/A N/A
Overall16 7.57.6 8.5 🥇

Quality evaluated by Claude Sonnet 4.6 (Anthropic, claude-sonnet-4-6) as independent AI evaluator. Criteria: factual accuracy, reasoning depth, specificity vs ground truth, output structure, hallucination absence.

What the responses actually showed — case highlights
CaseFindingWinner
text_02
Sheep reasoning
E4B: "There are 9 sheep left" — no explanation. E2B: "classic riddle." Qwen: step-by-step breakdown of the "all but = all except" idiom. Qwen
text_05
Meditation summary
Only Qwen mentioned cortisol levels — matching ground truth scientific specificity. E2B and E4B stayed generic. Qwen
img_02
Cyberpunk city
Only Qwen identified Chinese characters on neon signs, correctly placing the scene as East Asian. E2B missed the cyberpunk framing entirely. Qwen
img_04
Home library
Qwen hallucinated two fireplaces. E4B described the room accurately — single fireplace, correct layout. E4B
img_05
Neural network diagram
All 3 models called it a "network graph" or "social network." None identified it as a neural network architecture. Shared blind spot. None ✗
audio_01
Solar system
E2B: listed all 8 planets + "Jupiter is the largest planet" — near-perfect. E4B: "discusses the solar system and its planets" — missed all specifics. E2B
audio_03
Medical info
E4B produced structured bullet list with proper units (mmHg) — more clinically useful. E2B gave accurate prose. E4B
audio_04
Stanford AI news
E2B: perfect transcription. E4B: HTTP 500 — complete server failure. Only failure in the entire run. E2B
4 Key Findings

1 — Speed: E2B wins by 2.4×

RTX 3090 GDDR6X (936 GB/s) vs GB10 LPDDR5X (~273 GB/s). LLM inference is memory-bandwidth-bound at batch size 1. A 2020 GPU outpaces the 2025 DGX Spark for token generation.

2 — Quality: Qwen leads text + image

8.5 / 10 overall vs 7.6 (E4B) and 7.5 (E2B). More specific language, better reasoning explanations, stronger visual analysis. E4B edges out on hallucination control (library case).

3 — Audio: E2B is the only reliable choice

E2B scored 8.5 on audio: near-perfect transcription across all 5 cases. E4B scored 7.5 with 1 hard failure (HTTP 500 on audio_04). For production audio pipelines, E2B is the safe pick.

4 — Parameters beat precision at this scale

E4B (4-bit, 4B params) scores higher quality than E2B (bf16, 2B params) on image tasks. More parameters compensate for quantization. Don't assume full precision = better quality.

Deployment recommendation

Need speed? → Gemma4-E2B on RTX 3090 (47.5 tok/s, 100% reliable). Need best response quality for text/image? → Qwen3.5-4B on DGX (8.5/10). Need audio or video? → Gemma4-E2B is the only fully reliable multimodal option across all modalities.