Google launched Gemma 4 with four open models under Apache 2.0 license. The 31B model ranks #3 on Arena AI, beating models 20x its size. With native agentic workflows and edge-optimized variants, it's Google's strongest open-source play yet.
Article text
Google just made the open-source AI race more interesting. On April 2, 2026, the company released Gemma 4—a family of four open models that rank among the world's best while running on hardware that fits in your pocket.
The headline: the 31B model sits at #3 on Arena AI's text leaderboard. The 26B version ranks #6. Both are beating models with 20 times their parameter count.
But the real story isn't just performance. It's the Apache 2.0 license.
The License Shift That Matters
Previous Gemma models shipped under Google's custom Gemma license—not truly open-source in the conventional sense. Developers could use them, but with restrictions.
Gemma 4 changes that. Every model in the family ships under
Apache 2.0
, one of the most permissive open-source licenses available:
Commercial use
: Fully permitted without royalties
Modification
: Free to alter and distribute
Private use
: No reporting requirements
No copyleft
: Derivative works don't need to stay open
This puts Gemma 4 in direct competition with Meta's Llama models on licensing terms—something the open-source community has been asking for since Gemma's debut in February 2024.
Four Models, Four Use Cases
Gemma 4 isn't a single model. It's a family designed to scale from smartphones to data centers.
E2B (Effective 2 Billion Parameters)
Target
: Smartphones, IoT devices, Raspberry Pi, Jetson Orin Nano
Architecture
: Edge-optimized for battery and memory efficiency
Multimodal
: Vision + native audio input
Context
: 128K tokens
E4B (Effective 4 Billion Parameters)
Target
: Same edge devices, with more headroom
Trade-off
: Slightly more capable, slightly more resource-intensive
Multimodal
: Vision + native audio input
Context
: 128K tokens
26B Mixture of Experts (MoE)
Target
: PCs, consumer GPUs (quantized), workstations
Active parameters
: Only 3.8B during inference—hyper-optimized for latency
Specialty
: Fast token generation for responsive applications
Context
: Up to 256K tokens
Ranking
: #6 on Arena AI text leaderboard
31B Dense
Target
: Workstations, servers, H100 GPUs
Specialty
: Maximum quality, ideal foundation for fine-tuning
Context
: Up to 256K tokens
Ranking
: #3 on Arena AI text leaderboard
The efficiency story is striking. The 26B MoE activates only 3.8 billion parameters during inference, meaning you get near-flagship performance with a fraction of the compute cost.
Performance Benchmarks: Small but Mighty
The numbers don't lie. Gemma 4 is punching far above its weight class.
Benchmark
Gemma 4 31B
Gemma 4 26B MoE
Gemma 3 27B
Arena AI (Text)
1452
1441
1365
MMLU (Multilingual)
85.2%
82.6%
67.6%
AIME 2026 (Math)
89.2%
88.3%
20.8%
GPQA Diamond (Science)
84.3%
82.3%
42.4%
τ2-bench (Agents)
86.4%
85.5%
6.6%
The τ2-bench agent benchmark is particularly noteworthy. Gemma 4 scores 86.4% compared to Gemma 3's 6.6%. That's a 13x improvement in agentic capabilities.
Artificial Analysis, an independent benchmarking service, confirms the quality. On GPQA Diamond (scientific reasoning), Gemma 4 31B scores 85.7% in reasoning mode—second-best among all open models with fewer than 40 billion parameters.
Built for Agentic Workflows
Gemma 4 isn't just a chat model. It's designed for agents:
Function calling
: Native support for tool use
Structured output
: JSON formatting built-in
System instructions
: Define agent behavior and constraints
Multi-step reasoning
: Plan and execute complex workflows
This matters because the industry is shifting from chatbots to agents. Models that can't reliably call tools or follow structured instructions are falling behind.
Multimodal by Default
Every Gemma 4 model handles images and video. The E2B and E4B go further:
Native audio input
: Speech recognition built directly into the model
Near-zero latency
: Optimized for real-time voice interactions
Offline capable
: Run entirely on-device without internet
For mobile developers, this is significant. A single model that handles vision, audio, and text—without cloud dependencies—enables new categories of privacy-preserving applications.
The Developer Ecosystem
Google isn't releasing these models into a vacuum. The ecosystem support is comprehensive:
Inference frameworks:
Hugging Face Transformers
vLLM