Deploy OpenAI's open-weight reasoning models on your own hardware. The quality and behavior your teams already know from the OpenAI API — now with complete data sovereignty, no per-token cost, and no data leaving your servers.
# Same familiar OpenAI-style calls, # now served from your own cluster response = client.chat.completions.create( model="gpt-oss-120b", messages=[{ "role": "user", "content": "Plan a migration and call the tools" }], reasoning_effort="high", tools=agent_tools )
✓ Served on-prem — 0 bytes left your network
→ reasoning_effort=high · tool calls resolved locally
21B / 117B
Two MoE Model Sizes
128K
Context Window
Low/Med/High
Configurable Reasoning Effort
Apache 2.0
Open-Weight License
gpt-oss brings the model family teams already trust inside your own perimeter — no API keys leaving the building, no metered billing.
Weights run entirely within your environment. Prompts, documents, and outputs never leave your servers — deploy fully air-gapped if you need to.
Replace per-token cloud billing with fixed infrastructure cost. Run unlimited inference at a predictable rate — no rate limits, no surprise invoices.
Keep regulated data in-region and under your own controls. Ideal for finance, healthcare, government, and any workload with strict residency requirements.
gpt-oss is OpenAI's first open-weight release since GPT-2, licensed under Apache 2.0. Teams that already build on OpenAI's API get the behavior and prompting patterns they know — the difference is where it runs. Point your existing integrations at a deployment you own, and nothing changes for your developers except the base URL.
Pick the model that fits your hardware. Both are reasoning models with configurable reasoning effort and a 128K context window.
The single-GPU model. Around 21B total parameters with only ~3.6B active per token, so it stays fast and light while keeping strong reasoning.
Total / Active Params
~21B / ~3.6B
Minimum VRAM
~16 GB
Context Window
128K
Architecture
MoE
The flagship model that fits on a single 80GB GPU. Around 117B total parameters with only ~5.1B active per token — flagship reasoning without a multi-GPU rack.
Total / Active Params
~117B / ~5.1B
Fits On
1× 80GB GPU
Context Window
128K
Architecture
MoE
gpt-oss models are reasoning models — strong at agentic workflows, tool use, and chain-of-thought. Dial reasoning effort up or down to match the task.
Run multi-step agents that plan, act, and self-correct — entirely inside your network, with no external API calls.
Connect the model to your own systems and APIs. It reasons about which tools to call and chains them to complete real tasks.
Set reasoning effort to low, medium, or high per request — trade latency for depth on hard problems, keep it lean for simple ones.
We install gpt-oss on your infrastructure with an OpenAI-compatible endpoint. Your existing integrations keep working — you just change the base URL.
# Pull the deployment image
docker pull llmdeploy/gpt-oss:120b
# Or use our Python SDK
pip install sovereign-ai
# Initialize and run
from sovereign_ai import GptOss
model = GptOss("120b")
response = model.generate(
prompt="Draft a rollout plan for the audit",
reasoning_effort="high",
max_tokens=1000
)
gpt-oss-20b
~16GB VRAM · single GPU / workstation
gpt-oss-120b
Single 80GB GPU (e.g. one H100)
Deployment
Kubernetes, Docker, Bare Metal, Air-Gapped
Monitoring
Prometheus, Grafana, OpenTelemetry
Drop-in replacement for existing OpenAI integrations — familiar to any team already on the OpenAI API
# Simply change your base URL
client = OpenAI(
base_url="https://your-deployment.com/v1",
api_key="your-sovereign-ai-key"
)
72 hours
From kickoff to a running on-prem endpoint
Break-even
Fixed hardware cost overtakes metered cloud APIs at scale
0 bytes
Of your data leave your network — air-gap capable
Self-host gpt-oss with full data sovereignty, fixed cost, and the OpenAI-compatible workflow your team already knows.