Google just dropped Gemma 4 on April 2, 2026, and the benchmarks aren’t even the headline. Four model sizes. Native multimodal support. A 256K context window. But the real story? Google shipped it under Apache 2.0 — the first time the Gemma family has carried a genuinely open-source license. For anyone building enterprise AI infrastructure or evaluating sovereign deployment options, Google Gemma 4 just rewrote the playbook.
I’ve spent the last year deploying open models in production at a telecom. Licensing has killed more deployment plans than model quality ever has. Gemma 3’s custom license had carve-outs that made legal teams nervous. Llama 4’s community license still caps monthly active users. Gemma 4 under Apache 2.0 removes every one of those friction points — and the technical specs back up the ambition.
What Google Gemma 4 Actually Ships
Gemma 4 arrives in four variants, each available as both base and instruction-tuned models:
- Gemma 4 E2B — 2.3 billion active parameters (5.1B total). Native audio input. 128K context. Fits in under 1.5 GB quantized.
- Gemma 4 E4B — 4 billion effective parameters. Native audio input. 128K context. The sweet spot for on-device deployment.
- Gemma 4 26B MoE — Mixture of Experts architecture. Only 3.8 billion parameters active during inference from 26B total. 256K context window.
- Gemma 4 31B Dense — The flagship. Full dense architecture. 256K context. Ranked #3 globally among open models on LMArena with an ELO of approximately 1,452.
The “E” prefix on the smaller models stands for “effective parameters” — these use Per-Layer Embeddings (PLE), a technique that feeds a secondary embedding signal into every decoder layer. The result is that a 2.3B-active model carries the representational depth of a much larger network while fitting on a phone.
Every variant supports multimodal input out of the box: image understanding with variable aspect ratio, video comprehension up to 60 seconds (26B and 31B), and audio input for speech recognition and translation (E2B and E4B).
The Architecture Under the Hood
Three technical decisions separate Gemma 4 from its predecessors and competitors.
Alternating Attention
Gemma 4 layers alternate between local sliding-window attention and global full-context attention. Smaller dense models use 512-token sliding windows. Larger models use 1,024-token windows. The final layer is always global.
This is a deliberate engineering trade-off. Local sliding-window layers keep per-token compute linear in sequence length. Global layers — placed less frequently — handle long-range dependencies. The result: 256K context windows without the memory explosion that makes naive full-attention impractical at that scale.
Per-Layer Embeddings (PLE)
Standard transformers give each token a single embedding vector at input. PLE adds a parallel, lower-dimensional conditioning pathway alongside the main residual stream. For each token, it produces a small dedicated vector for every layer by combining a token-identity component with a context-aware component.
Each decoder layer then uses its corresponding PLE vector to modulate hidden states via a lightweight residual block. This is how the E2B model achieves the representational depth of a 5.1B parameter model while keeping active parameters at 2.3B — a meaningful advantage for edge deployment where every megabyte counts.
Dual RoPE
Standard rotary position embeddings for sliding-window layers. Proportional RoPE for global layers. This dual approach enables the 256K context window on larger models without the quality degradation that typically appears at long distances. It’s an elegant solution to a problem that has plagued long-context models for years.
The Apache 2.0 Licensing Shift Is the Real Story
When Google released Gemma 1, 2, and 3, each came with a custom license — the Gemma Terms of Use. That license imposed meaningful restrictions: limits on redistribution, specific attribution formats, and restrictions on applications exceeding certain monthly active user thresholds. Enterprise legal teams had to review every deployment.
Gemma 4 under Apache 2.0 changes the math completely:
- No custom clauses. No “Harmful Use” carve-outs requiring legal interpretation.
- No user limits. Deploy to 10 users or 10 billion. Same license.
- No redistribution restrictions. Embed it in commercial products, fork it, ship modified weights — all permitted.
- OSI-approved. This isn’t “open-ish” or “source available.” It’s genuinely open source by the Open Source Initiative’s definition.
Compare this to the competitive landscape. Llama 4’s community license still imposes a 700 million monthly active user threshold and includes an acceptable use policy that Meta can enforce. Qwen 3.5 ships under Apache 2.0 — but Gemma 4 now matches that openness while delivering competitive or superior benchmarks.
For enterprises evaluating sovereign AI deployments — running inference on-premises or in a private cloud — Apache 2.0 eliminates per-token API costs, provides full data sovereignty, removes rate limiting, and requires zero licensing overhead. That’s not an incremental improvement. It’s a different category of deployment.
Benchmarks: Where Gemma 4 Leads and Where It Doesn’t
The 31B Dense model posts strong numbers across the board:
| Benchmark | Gemma 4 31B | Notable Comparison |
|---|---|---|
| MMLU Pro | 85.2% | Exceeds Qwen 3.5 27B |
| AIME 2026 | 89.2% | Competitive with 70B+ models |
| Codeforces ELO | 2,150 | Competitive with much larger models |
| LMArena ELO | ~1,452 | #3 among open models globally |
The 26B MoE model ranks 6th on Arena AI — remarkable given that only 3.8B of its 26B parameters are active during inference. That efficiency ratio matters in production where you’re paying for compute per token.
Where Gemma 4 falls short: coding benchmarks. The broader Qwen 3.5 family leads on LiveCodeBench and SWE-bench with clear margins over both Llama 4 and Gemma 4. If your primary use case is code generation, Qwen still has the edge.
The E2B and E4B models occupy a space with no direct competition. Native audio support with 128K context in a sub-5B parameter model? Neither Llama 4 nor Qwen 3.5 offer anything comparable at that size tier.
NVIDIA Optimization: From Data Center to Your Desk
Google and NVIDIA collaborated on day-one optimization for Gemma 4 across the full NVIDIA stack. This isn’t a future promise — it’s shipping now:
- RTX PCs and Workstations — Run Gemma 4 locally via Ollama or llama.cpp with NVIDIA Tensor Core acceleration.
- DGX Spark — NVIDIA’s personal AI supercomputer handles the 31B model comfortably.
- Jetson Orin Nano — Edge deployment for the E2B and E4B models on IoT and embedded systems.
- Data Center GPUs — Full optimization across A100, H100, and B200 via the CUDA stack.
Unsloth provides day-one support with optimized quantized models for efficient local fine-tuning. The practical implication: you can fine-tune Gemma 4 on a single RTX 4090 for domain-specific tasks without touching a cloud API.
This matters because the cost structure of AI inference is shifting. Running a 26B MoE model locally with 3.8B active parameters on a $1,600 GPU eliminates the per-query cost entirely. For high-volume inference workloads — customer service, document processing, internal search — the ROI math favors local deployment over API calls within months.
What This Means for Enterprise AI Strategy
Three implications that matter if you’re making infrastructure decisions right now.
The Open-Source Tier Is Now Production-Grade
Gemma 4’s 31B model at #3 on LMArena sits in territory that was exclusive to proprietary APIs twelve months ago. The gap between open models and closed APIs has compressed to the point where the licensing and deployment flexibility of open weights outweighs the marginal quality advantage of frontier APIs for most enterprise use cases.
Sovereign AI Just Got Cheaper
Governments and regulated industries that need AI capabilities without data leaving their jurisdiction now have a genuinely open, high-quality option. Apache 2.0 means no licensing negotiation, no vendor lock-in, no per-seat fees. Google’s Sovereign Cloud offerings will support Gemma 4 across public cloud with data boundary, Google Cloud Dedicated, and air-gapped on-premises deployments.
The Three-Way Open-Source Race Intensifies
Gemma 4 vs Llama 4 vs Qwen 3.5 is now the defining competition in open AI. Each has different strengths — Llama 4 leads on context length (10M tokens), Qwen 3.5 leads on coding, and Gemma 4 leads on reasoning and multimodal at smaller sizes. The winner for any given deployment depends on the use case, not a single benchmark.
For practitioners building AI infrastructure, this three-way competition is unambiguously good news. Model quality is improving quarterly. Licensing is getting more permissive. Deployment tooling is maturing. The question is no longer “can open models handle production workloads?” — it’s “which open model fits this specific workload best?”
FAQ
What license does Google Gemma 4 use?
Gemma 4 is released under the Apache 2.0 license, an OSI-approved open-source license. This represents a major shift from previous Gemma versions, which used Google’s custom Gemma Terms of Use with restrictions on commercial deployment and user thresholds. Apache 2.0 grants irrevocable rights to use, modify, and distribute the models commercially with no royalty requirements.
How does Gemma 4 compare to Llama 4 and Qwen 3.5?
Gemma 4 31B leads in reasoning benchmarks (85.2% MMLU Pro, 89.2% AIME 2026) and ranks #3 among open models on LMArena. Llama 4 Scout offers a larger 10M token context window. Qwen 3.5 leads on coding benchmarks including LiveCodeBench and SWE-bench. The best choice depends on your specific use case — reasoning and multimodal favor Gemma 4, coding favors Qwen, and extreme context length favors Llama.
Can I run Gemma 4 locally on my own hardware?
Yes. NVIDIA has optimized Gemma 4 for local deployment from day one. The E2B model fits in under 1.5 GB quantized and runs on consumer hardware. The 26B MoE model activates only 3.8B parameters during inference, making it practical on a single RTX 4090. Tools like Ollama, llama.cpp, and Unsloth provide ready-to-use deployment and fine-tuning support.
What is Per-Layer Embeddings (PLE) in Gemma 4?
PLE is a technique used in Gemma 4’s smaller models (E2B and E4B) that feeds a secondary embedding signal into every decoder layer. Instead of the standard single embedding vector at input, PLE adds a parallel conditioning pathway that gives each layer its own token-specific modulation vector. This allows the E2B model to achieve the representational depth of a 5.1B parameter model while keeping active parameters at 2.3B.
Is Gemma 4 suitable for enterprise and government deployment?
Yes. The Apache 2.0 license removes all commercial restrictions, making Gemma 4 suitable for enterprise deployment without licensing overhead. Google supports Gemma 4 across Sovereign Cloud offerings including air-gapped on-premises deployments, which meets the data residency requirements of government and regulated industries. The model’s strong benchmark performance makes it competitive with proprietary APIs for most enterprise use cases.
What Comes Next
Google Gemma 4 isn’t just another model release. It’s a signal that the open-source AI tier has reached a quality threshold where the deployment and licensing advantages of open weights create a compelling alternative to proprietary APIs for the majority of enterprise workloads.
If you’re evaluating AI infrastructure decisions in Q2 2026, here’s the concrete next step: download the Gemma 4 26B MoE via Ollama, benchmark it against your current API-based workflows, and run the cost comparison. With 3.8B active parameters delivering top-6 global performance, the economics of local inference have never been more favorable. The gap between “good enough for production” and “frontier model quality” just got small enough that the licensing terms matter more than the benchmark delta.