← Back to feed

Google DeepMind released Gemma 4 on April 2, 2026 — a family of four open-weight models spanning edge devices to datacenter GPUs, all licensed under Apache 2.0.

Model Lineup

Variant	Params	Context	Hardware
E2B	2B	128K	Phone, Raspberry Pi, Jetson Nano
E4B	4B	128K	Phone, IoT
26B A4B (MoE)	26B total / 3.8B active	256K	16GB VRAM GPU
31B Dense	31B	256K	24GB VRAM (RTX 4090)

All four variants support multimodal input (image + video). Audio is available on the two edge models.

Key Benchmark Numbers

Benchmark	Gemma 4 31B	Gemma 3 27B	Delta
AIME 2026 Math	89.2%	20.8%	+68.4 pts
LiveCodeBench v6	80.0%	38.0%	+42.0 pts
GPQA Diamond	84.3%	50.0%	+34.3 pts
MMLU Pro	85.2%	67.5%	+17.7 pts
Chatbot Arena ELO	~1452	~1365	+87 pts

The 26B MoE activates only 3.8B parameters at inference, delivering near-31B quality at substantially lower compute cost. That design is the key architectural story: the MoE lands at Arena ELO ~1441, just 11 points below the dense 31B.

Where It Ranks

On Chatbot Arena as of April 2026, Gemma 4 31B sits third among all open models globally. It outperforms Llama 4 Maverick (ELO ~1417) and Qwen 3.5 27B on most benchmarks despite having fewer parameters than either.

Competitive programmers hit roughly ELO 2150 on Codeforces. Gemma 4 31B matches that threshold on the same benchmark.

Pricing

OpenRouter lists the 31B at $0.14/M input, $0.40/M output. At that price point it competes directly with Llama 4 Scout ($0.08/$0.30) while offering materially better reasoning.

Availability

Models are live on Hugging Face, Kaggle, Ollama, Google AI Studio, and AI Edge Gallery. The 26B and 31B run fully in AI Studio without local setup. The Apache 2.0 license allows commercial use and fine-tuning without royalties.

Gemma 4 shares its research base with Gemini 3 — same architecture lineage, different deployment target.