ArchRay
Build your on-prem AI architecture
Pick your ERP and documents, an open model, the GPUs, the retrieval pipeline and the guardrails. The diagram, memory, speed and cost update as you swap parts. Export it for your team, or have an engineer check it before you buy hardware.
Templates
- ERP copilot for one plant. Plain-English questions over live ERP data plus SOP search, about 150 users.
- Air-gapped, ITAR-style shop. No internet path at all; documents, drawings and ERP stay inside the enclave.
- Start small on one box. A pilot for 20 people on a single workstation GPU.
- Enterprise AI CoE, several models. A large model for reasoning, a small one for extraction, one gateway, 1,000 users.
- Document AI: POs, invoices, certs. Read supplier documents and post drafts to the ERP for approval.
- Text-to-SQL over JD Edwards. Ask F-table questions in English with the query shown, read-only.
Models
- Qwen3-VL-8B-Instruct on-prem
- Qwen3.5-9B on-prem
- Qwen3.8-27B on-prem
- Qwen3-30B-A3B-Instruct-2507 on-prem
- Qwen3-32B on-prem
- Qwen3.6-35B-A3B on-prem
- Qwen3-Coder-Next on-prem
- Qwen3.5-122B-A10B on-prem
- Qwen3.8-Flash-Next on-prem
- Qwen3.5-397B-A17B on-prem
- Qwen3.8-2.4T-A95B on-prem
- Olmo 3.1 32B Think on-prem
- Ling-3.0-flash on-prem
- Trinity Large Thinking on-prem
- Command A (03-2025) on-prem
- Command A+ (05-2026) on-prem
- DeepSeek-V3.2 on-prem
- DeepSeek-V4.1-Flash-1017 on-prem
- All 52 models
Hardware
- NVIDIA L4 24GB for LLMs
- NVIDIA A10 24GB for LLMs
- NVIDIA L40S 48GB for LLMs
- NVIDIA A100 40GB PCIe for LLMs
- NVIDIA A100 80GB PCIe for LLMs
- NVIDIA A100 80GB SXM for LLMs
- NVIDIA H100 PCIe 80GB for LLMs
- NVIDIA H100 SXM 80GB for LLMs
- NVIDIA H100 NVL 94GB for LLMs
- NVIDIA H200 SXM 141GB for LLMs
- NVIDIA H200 NVL 141GB for LLMs
- NVIDIA B200 180GB for LLMs
- NVIDIA B300 (Blackwell Ultra) 270GB for LLMs
- NVIDIA RTX A6000 48GB for LLMs
- NVIDIA RTX 6000 Ada 48GB for LLMs
- NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96GB for LLMs
- NVIDIA RTX PRO 6000 Blackwell Max-Q 96GB for LLMs
- NVIDIA RTX PRO 6000 Blackwell Server Edition 96GB for LLMs
- All 29 GPUs and systems
How it works
Every box in the diagram is a real, swappable part: 112 open-source and commercial components across serving engines, embedding models, rerankers, vector stores, document parsers, agent frameworks, gateways, guardrails, identity, observability and platforms. ERPs connect through the methods each vendor supports, such as IDO REST for SyteLine, RFC and OData for SAP, AIS for JD Edwards and the public APIs for Oracle EBS, shown read-only unless you choose otherwise.
Click any box to change it. Use Swap model and Swap hardware to compare every option on your own workload at once.
Questions
How does ArchRay size the GPUs?
Memory is the model weights at the precision you choose, plus the KV cache for every request generating at peak (computed from the model's layers, KV heads and head size, or its compressed latent for MLA models), plus runtime overhead. Speed is limited by memory bandwidth or compute, whichever is slower. It is a planning model, not a benchmark.
Which models and GPUs are included?
52 open-weight models and 29 GPUs and small systems, with architecture numbers taken from each model's published configuration and GPU numbers from vendor datasheets. We add new releases as they ship.
Can I share or export the design?
Yes. Download the diagram as PNG or SVG, the full configuration as JSON, or a bill of materials as CSV. The Share button copies a link that reopens the exact design.
Does anything I enter leave my browser?
No. The builder runs entirely in your browser. Only if you press Get it reviewed do we receive the design summary, so an engineer can reply.