Deploy Google DeepMind's open-weight models on your own hardware. Multimodal, multilingual, and engineered to run on a single GPU — strong on-premise AI without the cost and complexity of a multi-GPU cluster.
from sovereign_ai import Gemma3 # One GPU. Text + images. On your hardware. model = Gemma3("27B", device="cuda:0") response = model.generate( prompt="Summarize this scanned invoice", image="invoice_scan.png", max_tokens=1024 )
✓ Loaded Gemma 3 27B on 1× H100 — no cluster required
→ Multimodal request processed on-premise
1B–27B
Model Sizes
128K
Context Window
Single GPU
27B on One H100/A100
140+
Languages Supported
Gemma 3 scales from a text-only edge model to a multimodal flagship that fits on a single GPU — pick the size that matches your hardware, not the other way around.
Text-only · 32K context
• Runs on CPU / edge devices
• Lightweight assistants & classification
• Lowest footprint in the family
Multimodal · 128K context
• Text + image understanding
• Fits on consumer GPUs
• On-device & branch-office use
Multimodal · 128K context
• Text + image understanding
• Single mid-range GPU
• Balanced quality vs cost
Recommended — fits 1× H100/A100
• Multimodal · 128K context
• Highest quality in the family
• No multi-GPU cluster needed
Gemma 3 is designed for the best performance-per-parameter in its class. The 27B flagship runs on a single H100 or A100 — while the smaller sizes fit on consumer GPUs, or even CPU and edge hardware. That means the lowest-cost path to sovereign, on-premise AI: no multi-GPU cluster to buy, power, or maintain.
1× GPU
All the way up to the 27B flagship
Open weights, commercial use permitted under the Gemma terms — deployed entirely inside your network.
Every prompt, document, and image stays on your infrastructure — nothing leaves your network.
The 4B, 12B, and 27B sizes understand text and images together, across 140+ languages.
Pay for hardware once instead of per-token cloud API fees — with no usage caps or rate limits.
The efficient, single-GPU multimodal option — built for teams that want strong on-prem AI without a cluster.
Get sovereign, private AI without a six-figure GPU cluster. A single H100 or A100 runs the 27B flagship, and smaller sizes run on hardware you may already own — the most affordable route to on-prem AI.
The 1B text model runs on CPU and edge devices, while the 4B multimodal model fits on consumer GPUs — perfect for retail sites, factory floors, and branch offices where a datacenter isn't an option.
With multimodal support on the 4B, 12B, and 27B sizes, Gemma 3 reads scanned invoices, forms, diagrams, and photos alongside text — ideal for back-office automation, claims processing, and document intelligence entirely on-premise.
Support for 140+ languages and a 128K-token context window means one deployment serves teams and customers worldwide — long documents, multilingual support tickets, and cross-language search, all kept in-house.
From single GPU to production — with your existing tooling and OpenAI-compatible APIs.
# Install with Docker
docker pull llmdeploy/gemma3:27b
# Or use our Python SDK
pip install sovereign-ai
# Initialize and run on a single GPU
from sovereign_ai import Gemma3
model = Gemma3("27B", device="cuda:0")
response = model.generate(
prompt="Extract line items from this invoice",
image="invoice.png",
max_tokens=1000
)
Hardware
Single H100/A100, consumer GPUs, CPU/edge
Serving
vLLM, Ollama, TGI, llama.cpp
Deployment
Kubernetes, Docker, Bare Metal
Monitoring
Prometheus, Grafana, OpenTelemetry
Drop-in replacement for existing OpenAI integrations — swap the base URL and keep your code
# Simply change your base URL
client = OpenAI(
base_url="https://your-deployment.com/v1",
api_key="your-sovereign-ai-key"
)
72 hrs
From kickoff to production
1 GPU
Lowest hardware barrier to entry
Break-even
vs cloud APIs, often within months
Deploy Gemma 3 on a single GPU — multimodal, multilingual, and entirely on your own hardware.