Gemma 4 Complete Guide: Architecture, Models, and Deployment in 2026

Google DeepMind released Gemma 4 on April 3, 2026 under Apache 2.0 —
a significant licensing shift from previous Gemma releases that makes
it genuinely usable for commercial products without legal ambiguity.

This guide covers the full model family, architecture decisions worth
understanding, and practical deployment paths across cloud, local,
and mobile.

The Four Models and When to Use Each

Gemma 4 ships in four sizes with meaningfully different architectures:

Model	Params	Active	Architecture	VRAM (4-bit)	Target
E2B	~2.3B	all	Dense + PLE	~2GB	Mobile / edge
E4B	~4.5B	all	Dense + PLE	~3.6GB	Laptop / tablet
26B A4B	25.2B	3.8B	MoE	~16GB	Consumer GPU
31B	30.7B	all	Dense	~18GB	Workstation

The E2B result is the most surprising: multiple community benchmarks
confirm it outperforms Gemma 3 27B on several tasks despite being
12x smaller in effective parameter count.

Architecture: What's Actually Different

MoE vs Dense

The 26B A4B is a Mixture-of-Experts model. Despite 25.2B total
parameters, only 3.8B activate per token during inference. This means
self-hosting it requires significantly less VRAM than a comparable
dense model — closer to running a 4B model than a 26B one.

Gemma's MoE implementation differs from DeepSeek and Qwen: instead of
replacing MLP blocks with sparse experts, Gemma adds MoE blocks as
separate layers alongside the standard MLP blocks and sums their
outputs. This is an unusual design choice that trades some efficiency
for architectural simplicity.

Per-Layer Embeddings (PLE) in Edge Models

The E2B and E4B use PLE instead of MoE — a different efficiency
strategy suited for mobile inference. Standard transformers give each
token a single embedding vector at input. PLE adds a parallel
lower-dimensional conditioning pathway: for each token, it produces
a small dedicated vector per layer, letting each decoder layer receive
token-specific information only when relevant rather than requiring
everything to be frontloaded into a single embedding.

This is what enables E2B to run under 1.5GB RAM on supported mobile
devices via LiteRT-LM.

Hybrid Attention

All Gemma 4 models use alternating local sliding-window and global
full-context attention layers. Smaller models use 512-token sliding
windows, larger ones use 1024. The final layer is always global.

For KV cache optimization, global layers share key-value states from
earlier layers (Shared KV Cache), eliminating redundant KV projections.

Known issue: The KV cache footprint at long context is substantial.
Community reports indicate the 31B at 262K context requires ~22GB just
for KV cache on top of the model weight. Workaround:
--ctx-size 8192 --cache-type-k q4_0 --parallel 1

Context Windows and Multimodal Capabilities

Feature	E2B / E4B	26B A4B / 31B
Context	128K	256K
Image input	✅	✅
Video input	✅ (60s @ 1fps)	✅
Audio input	✅	❌
Function calling	✅	✅

Audio input is edge-model only — E2B and E4B support ASR and
speech-to-translated-text via a USM-style conformer encoder.

Local Deployment

Ollama (fastest to get running)

# E4B — recommended for most laptops
ollama pull gemma4:e4b
ollama run gemma4:e4b

# 26B A4B — needs 16GB+ VRAM
ollama pull gemma4:26b-a4b

Note: requires Ollama 0.20 or newer for Gemma 4 support.

llama.cpp

A tokenizer fix was merged into the main branch shortly after launch.
Pull the latest and recompile before running Gemma 4 GGUF files.

Apple Silicon (MLX)

Unsloth MLX builds use ~40% less memory than Ollama at the cost of
~15-20% lower token throughput. For memory-constrained setups:

pip install mlx-lm
mlx_lm.generate --model unsloth/gemma-4-e4b-it-mlx --prompt "Hello"

LM Studio

Search "gemma4" in the model browser. E4B and 26B A4B are available
as pre-quantized GGUF files.

Mobile Deployment

Android

Android has the most complete official on-device story:

1. Google AI Edge Gallery — install from Play Store, select Gemma 4 E2B or E4B, runs fully on-device
2. LiteRT-LM — for developers building their own apps, gets E2B under 1.5GB RAM with 2-bit and 4-bit quantization
3. ML Kit GenAI Prompt API — production-ready API for Android app integration
4. Android AICore — system-wide optimized Gemma 4 access on supported Android 10+ devices

iOS

iOS is currently a developer integration story, not a consumer one.
The official path is MediaPipe LLM Inference SDK. No App Store
consumer app with Gemma 4 exists yet.

A practical reference for both Android and iOS deployment paths
is available at gemma4.app/mobile.

Cloud Deployment

Google offers three official cloud paths:

Vertex AI — managed deployment with autoscaling, best for
production workloads requiring SLA guarantees.

Cloud Run — serverless container deployment, lower operational
overhead, suitable for moderate traffic.

Google Kubernetes Engine (GKE) — vLLM on GKE for high-throughput
serving, best for teams already running Kubernetes infrastructure.

For API access without self-hosting, the 26B A4B is available via
OpenRouter at $0.13/M input tokens and $0.40/M output tokens.

Benchmark Context

The 31B dense model ranks #3 among open models on Arena AI as of
launch. Key numbers:

• AIME 2026: 89.2% (31B with reasoning)
• GPQA Diamond: 85.7%
• LiveCodeBench v6: 80.0%

The ELO gap vs automated benchmarks is notable: the 31B scores
higher on human preference rankings (Arena) than raw benchmark
comparisons with Qwen 3.5 27B would suggest, indicating the model
produces outputs humans prefer even when accuracy is similar.

Fine-tuning Status

QLoRA fine-tuning tooling was not ready at launch. Three issues
were reported within the first 24 hours:

• HuggingFace Transformers didn't recognize the gemma4 architecture (required installing from source initially)
• PEFT couldn't handle Gemma4ClippableLinear, a new layer type in the vision encoder
• A new mm_token_type_ids field is required during training even for text-only data

Both huggingface/peft and huggingface/transformers issues have
been filed. Check repo status before attempting fine-tuning.

Summary

Gemma 4's practical value proposition by use case:

• Mobile / privacy-first apps → E2B or E4B via LiteRT-LM
• Local assistant on laptop → E4B via Ollama
• Best open model on consumer GPU → 26B A4B (MoE efficiency)
• Maximum quality, workstation → 31B dense
• Production cloud → Vertex AI or GKE with vLLM

The Apache 2.0 license removes the previous ambiguity for commercial
use. For teams evaluating open models for production deployment,
Gemma 4 is now a first-class option alongside Qwen 3.5 and Llama.