One command ships your whole canvas onto a GPU pod. DISPATCH provisions an A100 80GB on RunPod, serves a 70B-class model, and wires the endpoint back into every node — no containers, no inference server to stand up.
A DISPATCH workflow is already a self-contained graph — nodes, bindings, tools and edges in one file. Rapid Pod Deployments lifts that file onto a GPU pod, serves a 70B-class model on it, and repoints every node at the pod's local endpoint. No Dockerfile, no Kubernetes.
From local canvas to served pod.
DISPATCH snapshots the graph, requests the GPU, boots an inference server, and rebinds the nodes. The chain you tested on a hosted model now runs against one you serve yourself — on hardware big enough to hold it.
A100 80GB, 70B-class served.
Big models, no rig.
A 70B-class model won't fit on a laptop. The pod gives every node — Investigate, Plan, Create — a frontier-scale model without buying a GPU.
Co-located, low latency.
Orchestration and weights sit on the same pod, so a 20-wide Parallel fan-out talks to the model over loopback — no public round-trip per token.
Ephemeral by default.
Spin the pod up for a run, tear it down after. You pay for GPU-minutes, not a standing cluster — and the canvas file is the only thing you keep.
A live run of dispatch deploy --pod — canvas packaged, GPU acquired, weights loaded into VRAM, endpoint bound back to the nodes.
DISPATCH treats the pod as a managed target. Standing up a serving stack used to be a day of infra — here it's one line.
Stand up a serving stack.
Ship the canvas.
Ollama, llama.cpp, LM Studio — fully local, alongside the cloud.
openSet provider + model per node and swap live, mid-run.
openDeploy 20 agents on one task — parallel, streamed, synthesized.
open