Gemma 4: open MoE models from Google up to 31B parameters

Gemma 4 from Google DeepMind: New Open Models with MoE Architecture and Multimodality

Google DeepMind has introduced Gemma 4 — the next generation of open AI models based on Gemini 3. The lineup includes four variants: E2B, E4B, 26B MoE, and 31B Dense. The key novelty is the 26B model with Mixture of Experts (MoE) architecture, where only 3.8 billion parameters are activated during inference. This delivers high performance while preserving depth of knowledge. The flagship 31B Dense model took third place in the Arena AI Text leaderboard among open counterparts, showcasing efficiency in terms of "intelligence per parameter".

All models are multimodal: they process images and video. The compact E2B and E4B also support audio input for local speech recognition. Context window ranges from 128K tokens for the smaller models to 256K for the larger ones. Native support for function calling and structured JSON output is built in for agentic applications.

Architectural Features and Optimizations

The MoE in the 26B model minimizes computational load: out of 26 billion parameters, only 3.8 billion are actively used. This is crucial for low-latency tasks. The 31B Dense retains a dense architecture but is optimized to compete with models 2+ times larger.

The compact E2B and E4B are targeted at edge devices:

• Android smartphones;
• Raspberry Pi;
• Jetson Nano.

They run fully offline, without cloud dependencies. Support for 140+ languages and an Apache 2.0 license simplify commercial use without strict restrictions.

Model weights are available on Hugging Face, Kaggle, Ollama, LM Studio, and in Docker images. The larger models fit on a single GPU, which is relevant for scenarios requiring data privacy and digital sovereignty.

Capabilities for Developers

Gemma 4 is tailored for real-world tasks of mid/senior developers:

• Multimodality: input data — text, images, video, audio (for E2B/E4B).
• Agentic features: native function calling and structured JSON output without post-processing.
• Long context: up to 256K tokens for handling large documents or conversations.
• Local deployment: from mobile devices to servers with a single GPU.
• Multilingualism: 140+ languages without fine-tuning.

This enables integration into low-latency applications: chatbots, RAG systems, on-device AI.

Benchmarks and Comparisons

The 31B Dense outperforms many open models twice its size in Arena AI Text (3rd place). Google developers focused on efficiency: high metrics at a smaller size. Compact versions maintain performance for edge tasks where resources are limited.

For testing, use Hugging Face Transformers or Ollama. Example of basic 31B loading:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "google/gemma-4-31b"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float16, device_map="auto")

Inference on a single GPU with quantization (e.g., 4-bit) reduces memory to 20-25 GB.

Key Takeaways

• MoE optimization: 26B activates 3.8B parameters — balance of speed and quality.
• Edge support: E2B/E4B for Android, Raspberry Pi, offline speech.
• Agentic features: function calling + JSON out of the box.
• Accessibility: Apache 2.0, weights on HF/Ollama, 140+ languages.
• Efficiency: 31B competes with 60B+ models in benchmarks.

The Gemma 4 family expands local AI capabilities for production tasks without compromising functionality.

— Editorial Team