ai.canshu.com
中文
VRAM Calculator Advisor & TCO API Pricing Compare Models Leaderboard GPUs

Search models, GPUs, and tools

↑↓ Navigate · Enter Open · ★ Watchlist prioritized ai.canshu.com
Google DeepMind
Open-Weight Google DeepMind · MoE 8B / A4.5B · Text+Image+Video+Audio · BF16

Gemma-4-E4B-IT (Edge Multimodal) Architecture Parameters, Benchmarks & VRAM

On-device native text+image+audio model (~8B total with PLE tables, 4.5B effective active parameters).

🧬 Gemma-4-E4B-IT (Edge Multimodal) Architecture & Tensor Specifications

Verified directly against HuggingFace (google/gemma-4-E4B-it) config.json & Model Card

HF Class: Gemma4ForConditionalGeneration

1. Backbone & Dimensions

gemma4
Total / Active Params
8B / 4.5B
Non-Embedding Params
4.5B
Hidden Size (d_model)
2,560
FFN Intermediate Size
10,240
Activation / Norm
GELU-Tanh (GeGLU) · ε=1e-6
Native Tensor Dtype
BF16

2. Attention & Layer Layout

HYBRID
Num Hidden Layers
42 Layers (7 Global KV)
Hidden Layer Layout
42-Layer On-Device Omni Backbone (w/ Per-Layer Embeddings): 35 Sliding + 7 Global Layers
Attention Heads (Q / KV)
8 Q / 2 KV (4:1 GQA)
Head Dimension
512
Linear / KV Compression
Native Vision + Audio encoders (supports ASR & speech-to-text translation)
Vision Encoder / MTP
ViT 16L

3. Context, RoPE & Tokenizer

Tokenizer
Native Context Length
131,072 tokens
RoPE Configuration
Dual RoPE (Global θ=1,000,000 · Sliding θ=10,000 · 128K Context)
Vocab Size
262,144 (Tied)
Special Token IDs
BOS: 2 · EOS: [1,106,50] · PAD: 0
Language Support
140+ languages (including multilingual audio ASR)

🏆 Official Benchmark Results

Sourced from official System Card / Model Card & technical reports

Arena Elo ~1342 AA Index 29
📚 Knowledge
MMLU-Pro 69.4 %

A more robust and challenging multi-task language understanding benchmark that extends MMLU by expanding multiple-choice options from 4 to 10, eliminating trivial questions, and focusing on reasoning-intensive tasks.

📚 Knowledge
MMMLU 76.6 %

Multilingual Massive Multitask Language Understanding dataset released by OpenAI, featuring professionally translated MMLU test questions across 14 languages including Arabic, Bengali, German, Spanish, French, Hindi, Indonesian, Italian, Japanese, Korean, Portuguese, Swahili…

🧠 Reasoning
GPQA Diamond 58.6 %

PhD-level Google-proof science reasoning questions

🧠 Reasoning
AIME 2026 (no tools) 42.5 %
💻 Coding
LiveCodeBench v6 52 %

LiveCodeBench is a holistic and contamination-free evaluation benchmark for large language models for code.

👁️ Multimodal
MATH-Vision 59.5 %
👁️ Multimodal
MMMU-Pro 52.6 %

A more robust multi-discipline multimodal understanding benchmark that enhances MMMU through a three-step process: filtering text-only answerable questions, augmenting candidate options, and introducing vision-only input settings.

🎛️ Official Sampling & Reasoning Parameters

Recommended generation settings & API parameter constraints

Official Default Multimodal & Thinking
temperature=1 top_p=0.95 top_k=64 max_tokens=32,768

💡 Supports interleaved multimodal input, built-in thinking mode, and native function calling.

📊 Gemma-4-E4B-IT (Edge Multimodal) Full Benchmark Scorecard

12 public benchmarks · ranks among 154 models · synced 2026-11-20

⚔️ Compare head-to-head

Source: ZeroEval / llm-stats.com (✓ = third-party verified, others vendor-reported).

⚡ Gemma-4-E4B-IT (Edge Multimodal) Interactive VRAM & Token Speed Calculator

Concurrency (Batch):
MoE 8B / A4.5B·42L (d=2560)·8Q/2KV·Ctx 128K·Text+Image+Video+Audio
Runnable Models →
GPUs
32 GB VRAM·1792 GB/s·Safe Usable 31.4 GB
Smooth Fit6.25 / 31.4 GB~420 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
6.25/ 31.4 GB usable

Physical 32 GB · Load 19.9%

Theoretical Decode Speed
~420tokens/s

MoE: reads only 2.73 GB active weights/step

Max Context at Batch=1
128K/ 128K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 4.85 GBKV Cache (1×16K): 0.51 GBOverhead: 0.9 GB
🚀 Blazing Fast (Instant CoT)~1.2s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run gemma4:e4b

RTX 5090 Fits Gemma-4-E4B-IT (Edge Multimodal) · Cost & TCO

Approx. $2,199 (~$69/GB VRAM) · 575W TDP

📊 Gemma-4-E4B-IT (Edge Multimodal) Quantization VRAM Breakdown (16K Context)

Quantization Bits/Weight Weights KV Cache (16K) Total VRAM Quality Note
FP16 / BF16 16 bpw 16 GB 0.51 GB 17.63 GB Lossless Full Precision
FP8 / Q8_0 8.5 bpw 8.5 GB 0.51 GB 9.98 GB 99.9% Near-Lossless
GGUF Q6_K 6.56 bpw 6.56 GB 0.51 GB 8 GB Ultra High Quality
GGUF Q5_K_M 5.69 bpw 5.69 GB 0.51 GB 7.11 GB High Balanced Quality
Q4_K_M / AWQ 4.85 bpw 4.85 GB 0.51 GB 6.25 GB Sweet Spot (Recommended)
GGUF IQ4_XS 4.25 bpw 4.25 GB 0.51 GB 5.64 GB Compact 4-bit
GGUF Q3_K_M 3.91 bpw 3.91 GB 0.51 GB 5.29 GB Moderate Perplexity Loss
GGUF Q2_K 3.15 bpw 3.15 GB 0.51 GB 4.52 GB Extreme Compression

🖥️ GPU & Apple Silicon Compatibility for Gemma-4-E4B-IT (Edge Multimodal) (Q4_K_M)

Frequently Asked Questions

What are the exact architecture parameters of Gemma-4-E4B-IT (Edge Multimodal)?

Gemma-4-E4B-IT (Edge Multimodal) uses Gemma4ForConditionalGeneration with 42 layers (42-Layer On-Device Omni Backbone (w/ Per-Layer Embeddings): 35 Sliding + 7 Global Layers), hidden_size=2560, 8 Q heads / 2 KV heads (head_dim=512), and vocab_size=262,144.

How much VRAM does Gemma-4-E4B-IT (Edge Multimodal) require?

At 4-bit quantization (Q4_K_M, 16K context), Gemma-4-E4B-IT (Edge Multimodal) requires 6.25 GB VRAM. Full FP16 requires 17.63 GB.