One grid. Three deployment postures.
Deployment is a posture, not an identity. Start wherever fits your life today and move whenever that changes. Your agents, your packs, and your approval trail come with you.
Runs where you choose.
We host, you decide.
ArcGrid's managed infrastructure. Start this week, no hardware required. Your data lives in our multi-tenant isolated environment. Move to self-hosted whenever you choose.
Your hardware, your walls.
The full grid runs on hardware you own or control. Nothing leaves your network without your explicit action. ArcRig-class hardware runs the inference layer locally.
Local core, measured cloud.
Sensitive operations run on your hardware. Burst capacity and updates draw from managed cloud. You define the boundary. Switch the boundary anytime.
The stack, named honestly.
ArcGrid runs on production-grade open infrastructure. No proprietary lock-in at the infrastructure layer.
Local inference on hardware you control.
The ArcRig fleet is the inference backbone for self-hosted and hybrid deployments. Dedicated GPU/NPU hardware runs local models at production throughput. No cloud dependency for inference in self-hosted mode.
llama3.1:8b - fast classification, quick chat, routing decisions. Low-latency tier for high-frequency tasks.
qwen2.5-coder:32b + qwen3-fast. Code generation, structured analysis, long-form synthesis. Primary local workhorse.
For tasks that require frontier capability, the LiteLLM proxy routes to Anthropic Claude under full audit logging. All passthrough traffic is logged for training data.