The old Gemma license had a clause that let Google limit access to your deployed model "remotely or otherwise" if your usage was deemed a violation. Under the previous framework, Google retained the ability to modify usage terms at any time, and its Prohibited Use Policy was incorporated by reference. If a user was deemed to violate policies or regulations, Google could limit access "remotely or otherwise." That is not a theoretical risk most legal teams can accept. On April 2, 2026, it went away.
Gemma 4 is Google DeepMind's fourth-generation open-weights language model family, released on April 2, 2026 under the Apache 2.0 license with no commercial restrictions. That single licensing change is more significant than any benchmark number in the release notes - and it arrives in a model family that actually fits the hardware your team already owns.
What the Apache 2.0 switch actually removes
Apache 2.0 is not just a softer version of the old Gemma terms. Previous Gemma releases shipped with a custom "Gemma Terms of Use" license that placed restrictions on commercial deployment, redistribution, and derivative works, creating friction for teams building products on top of Gemma models. Apache 2.0 removes all of it - you can fine-tune Gemma 4, deploy it commercially, redistribute modified weights, and build closed-source products on top of it with no usage restrictions.
The comparison with Llama 4 is concrete: Llama 4's Community License restricts apps with over 700 million monthly active users and requires "Built with Llama" branding, creating potential legal risk for large enterprises. Gemma 4 and Qwen 3.5 are both Apache 2.0 - no such issues.
One thing the license does not give you: if you are in healthcare, legal, or finance, you must build your own audit trails and compliance monitoring - Google does not provide SOC 2 certification for self-hosted Gemma 4. That is not a dealbreaker, but it is work your team has to own.
Which model size fits which hardware
The Gemma 4 family spans four sizes: E2B (2.3B effective parameters), E4B (4.5B), 26B MoE (only 3.8B active per token), and 31B Dense - each capable of running entirely on consumer hardware.
The E-series uses Per-Layer Embeddings to keep the memory footprint small without proportionally cutting quality.
At INT4 quantization (Q4_K_M), the E2B needs only ~1.5 GB VRAM, the E4B ~3 GB, the 26B MoE ~16 GB, and the 31B Dense ~18 GB. INT8 roughly doubles these numbers.
| Model | Effective params | VRAM at INT4 | Fits on |
|---|---|---|---|
| E2B | ~2.3B | ~1.5 GB | Phone, Raspberry Pi 5 |
| E4B | ~4.5B | ~3 GB | Any laptop with 4 GB+ VRAM |
| 26B MoE | 3.8B active | ~16 GB | RTX 4090 (24 GB), M3 Max (48 GB) |
| 31B Dense | 31B | ~18 GB | RTX 4090, RTX 5090, M3 Max |
Key capabilities across the family include built-in reasoning mode that lets the model think step-by-step before answering, and context windows of up to 128K tokens (E2B/E4B) and 256K tokens (12B/26B/31B).
A few honest caveats: a 7B model that runs at 45 tokens/sec on GPU will drop to 8 tokens/sec the moment even 10% of its layers hit system RAM. Size the hardware to hold the full model in VRAM - offloading to system RAM is not a half-measure, it is a near-stop. And for multi-turn agent workloads specifically, CPU offloading is especially damaging because the KV cache grows with each turn, pushing you further into CPU territory even if the base model originally fit.
The economics of running it yourself
Self-hosting makes financial sense at a specific volume. Via API, Gemma 4 31B costs $0.12-$0.14 per million input tokens and $0.35 per million output tokens depending on provider. For high-volume use, self-hosting is dramatically cheaper: a $500 RTX 4090 GPU amortizes to roughly $0.001 per million tokens if you run 100M tokens per month. For lower volume (under 10M tokens/month), API providers make financial sense.
Do the math for a team running a private internal assistant: at 10 queries per person per day, 50 people, averaging 1,000 tokens per query, you hit about 150M tokens a month. At API pricing that is $21-$52.50 a month in output costs alone. On self-hosted hardware, the same volume costs the electricity bill and a fraction of the GPU's amortized cost. The crossover point is somewhere around 20-30M tokens/month on modest hardware.
For sensitive data or strict compliance requirements, self-hosting is often the only acceptable path
- the economics are almost secondary for regulated industries.
Beagle in action#legal-ops, 11:07amWhat Gemma 4 does not fix
Gemma 4 scores well on reasoning and math (AIME 2026 ~89%) and code generation (LiveCodeBench v6 ~80%), but it trails Qwen 3.5 27B on SWE-bench Verified, has no native speech output, and fine-tuning, weights, and serving become your problem.
The context window gap with Llama 4 Scout is real. Llama 4 Scout offers a 10M token context (Gemma 4 offers 256K) but requires 24 GB+ VRAM even quantized; Gemma 4 is lighter and runs on more hardware. For whole-codebase analysis or very long document workflows, Scout still has the edge. For most team use cases - summarization, classification, Q&A against internal docs, draft generation - 256K is enough.
The tooling ecosystem is also younger than Llama's. Community tooling maturity is a real consideration: Qwen and Llama have larger fine-tuning ecosystems. You can fine-tune Gemma 4 with LoRA on an RTX 3090 for the E4B, but the community of pre-built adapters and domain-specific checkpoints you'd find for Llama simply does not exist yet at the same depth.
Run a private open-weight AI model: common questions
What does it actually cost to run Gemma 4 privately?
A $500 RTX 4090 GPU amortizes to roughly $0.001 per million tokens at 100M tokens per month, versus $0.12-$0.35 per million tokens via API. Below 10M tokens per month, API access is likely cheaper when you factor in hardware and maintenance overhead.
Is Gemma 4 truly open-source? Can I use it commercially?
Apache 2.0 removes all of the restrictions from the prior Gemma Terms of Use. You can fine-tune Gemma 4, deploy it commercially, redistribute modified weights, and build closed-source products on top of it with no usage restrictions. There is no revenue threshold, no MAU cap, and no branding requirement.
Which Gemma 4 size should a team start with?
For most teams, the E4B (~3 GB at INT4) is the right entry point - it fits on any development machine with 4 GB+ VRAM and handles summarization, classification, and Q&A work reliably. Move to the 26B MoE when you need stronger reasoning or longer context and have a machine with a 16-24 GB GPU available.
How does Gemma 4 compare to Llama 4 for private deployment?
Llama 4's Community License restricts apps with over 700 million monthly active users and requires "Built with Llama" branding, creating potential legal risk for large enterprises. Gemma 4 is Apache 2.0 with no such restrictions. Llama 4 Scout's 10M context window beats Gemma 4's 256K for whole-codebase tasks, but Gemma 4 runs on smaller hardware.
Can you self-host a private AI model without a data center?
Yes. The minimum requirements to run a capable local LLM are 16 GB of system RAM, a modern CPU, and either a GPU with 6+ GB of VRAM or an Apple Silicon Mac - enough for a 3B-7B model at Q4. For the Gemma 4 E4B specifically, any machine with 4 GB of VRAM handles it comfortably at INT4.