Docs · Bring your own model

Your weights. Our fleet.

Any open-weights checkpoint, one spec file, OpenAI-compatible endpoint out. Custom Docker servers supported for exotic stacks.

runaii.yaml
# runaii.yaml — model packaging spec (Truss-inspired, minimal)
model: Qwen/Qwen3-32B
engine: vllm                      # vllm | sglang | tgi | custom
quant: nvfp4                      # fp8 | nvfp4 | bf16 | awq | gptq
context: 131072
gpus:
  min: 1
  max: 8
  types: [b200, b300]             # scheduler picks cheapest fitting
env:
  TENSOR_PARALLEL: "auto"
endpoints:
  - /v1/chat/completions         # OpenAI-compatible, always exposed
healthcheck:
  path: /health
  ready_contains: "ok"

1. Package

Write runaii.yaml next to your weights (HF repo, GCS, or upload). We validate the spec in CI before a single GPU spins.

2. Deploy to staging

One command or console click. Staging gets its own endpoint for evals — priced at standard dedicated rates.

3. Promote

Traffic-split promotion with automatic rollback on error-rate or latency SLO breach. Zero downtime by construction.

Engine support

vLLM and SGLang first-class (tensor + pipeline parallel auto-configured). TGI supported. Custom servers: expose /health and OpenAI-schema endpoints on port 8000 — we handle TLS, autoscaling, and metering.