01
Familiar integration
Point your preferred HTTP client or SDK at one base URL. Chat, streaming, tools, vision and embeddings follow familiar request formats.
base_url="https://llm.grand-ice.com/v1"
Private AI infrastructure
Stop renting intelligence by the token.
A self-hosted AI platform for teams that are done watching usage meters drain their technology budget. Run capable open models behind one secure, provider-grade API.
0per-token vendor fees
6purpose-built model options
100%self-hosted inference
POST /v1/chat/completions
Live
$ curl https://llm.grand-ice.com\
-H "Authorization: Bearer $KEY" \
-d '{
"model": "smart",
"messages": [...]
}'
-H "Authorization: Bearer $KEY" \
-d '{
"model": "smart",
"messages": [...]
}'
200 OKPrivate inference completed
2.4s
Model routesmart → qwen3.5:9b
Data destinationYour infrastructure
The economics are broken
Your AI budget should fund products—not unpredictable token bills.
Prototypes are cheap. Production is not. Once customers, agents and internal workflows start making millions of calls, every useful interaction becomes another line item.
Grandice changes the unit economics: invest in infrastructure once, then use it continuously without a vendor charging for every input and output token.
The traditional AI bill
Input tokensmetered
Output tokensmetered
Embedding callsmetered
Usage growthbudget risk ↑
Grandice token bill$0.00
Hardware, electricity and operations remain your responsibility.
One private AI layer
The capabilities teams expect from leading LLM platforms. The infrastructure you control.
Move from hosted providers without rebuilding your application stack.
02
Private by design
Prompts and model inference stay on infrastructure you operate—not in a third-party model provider’s cloud.
03
Predictable economics
No per-token vendor invoice. Capacity is determined by your hardware, concurrency and operating policy.
04
Governed access
Issue a unique key per team or application, restrict model access, set rate limits and review usage centrally.
05
Models without lock-in
Stable aliases decouple applications from model versions. Upgrade the model behind smart without touching client code.
06
Auditable domain workflows
Typed financial, legal, and market APIs separate deterministic analysis, caller-source metadata, privacy checks and professional-review requirements.
POST /v1/market/research/analyze
Read the financial agents white paper →
Read the legal contract intelligence white paper →
Read the market research intelligence white paper →
Grandice Privacy Shield
Sensitive data can stop before the network starts.
Grandice provides auditable, source-grounded domain workflows through developer-friendly APIs. Privacy Shield adds a local protection boundary for prompts, structured data and multimodal inputs before those workflows run.
Explore the privacy APIs →Client-side SDKs
Python and TypeScript protection runs inside your application. Original values and reversible token maps stay in your process.
mode: "client"
Local privacy sidecar
Point an existing OpenAI-compatible application at a loopback proxy for policy enforcement, local media processing and safe rehydration.
localhost:8090/v1
Hosted redaction API
Detect, mask or tokenize JSON with a single authenticated request. Original content reaches Grandice memory but is not written by the privacy layer.
POST /v1/privacy/redact
Text + JSONRecursive detection, masking and tokenization
Images + audio + videoLocal OCR, visual, speech and metadata pipelines
Fail closedUnsupported inputs are blocked instead of sent unprotected
Privacy receiptsPolicy and counts without original sensitive values
Local media runtime active
Multimodal protection is installed and validated on this host.
- Tesseract OCR
- Face detection
- QR + barcode detection
- FFmpeg frame processing
- Local Whisper transcription
- Metadata removal
Zero-transfer applies only when protection runs in the client SDK or local sidecar. Hosted processing is zero-retention, not zero-transfer. Automated detection has limitations and high-risk outputs still require review.
The model fleet
Choose the right intelligence for every workload.
Simple aliases make capability selection clear for developers and end users.
View model documentation →S
smartComplex work, vision, tools & code
Qwen 3.5 · 9B
F
fastLow-latency chat & vision
Qwen 3.5 · 4B
C
chatGeneral conversation & tools
Qwen 3 · 8B
</>
codeGeneration, review & completion
Qwen 2.5 Coder · 7B
◉
visionImages, screenshots & OCR
Qwen 2.5 VL · 7B
E
embedRAG & semantic search
Nomic Embed Text
Clear by architecture
Your apps connect securely. Your data stays close.
Client SDKs or the local sidecar can sanitize sensitive values before transmission. Cloudflare provides encrypted ingress, the authenticated gateway governs every request, and model weights remain isolated on the private host.
01Your applicationsSDK · local privacy
→
02Grandice gatewayAuth · policy · logs
→
03Local modelsPrivate inference
Built for responsible access
Freedom from token fees does not mean freedom from controls.
01
Per-application API keys
Separate credentials make access revocable and usage attributable.
02
Model-level permissions
Limit each key to the capabilities that application actually needs.
03
Rate limiting
Protect finite local capacity from accidental loops and abusive traffic.
04
Private model host
Ollama stays bound to localhost; only the authenticated gateway is published.
05
Privacy policy enforcement
Versioned policies, residual checks and fail-closed processing prevent known sensitive data from silently bypassing protection.
Build without the meter running