OhhO Serve
Robot AI inference, as an API.
- Category: Intelligence
- Accent: cyan
- Live app: Open the console
Deploy a Vision-Language-Action model behind a REST endpoint in one command. Pluggable model backends, batching and Prometheus metrics built in.
Overview (hero)
Robot AI inference, as an API. Deploy a Vision-Language-Action model behind a REST endpoint in one command — pluggable model backends, batching, and Prometheus metrics built in.
Highlights
- One-command model deploy
- REST predict API
- Pluggable backends (OpenVLA, SmolVLA, ACT…)
- 4-bit quantization option
- Prometheus metrics built in
What you get
- Modern robots run on large Vision-Language-Action (VLA) models, but getting one into production means GPU servers, model loading, batching, versioning and monitoring. OhhO Serve turns all of that into a single endpoint.
- Point Serve at a model — OpenVLA, SmolVLA, ACT, a diffusion policy, or your own fine-tune — and it exposes a clean REST API your robot calls with an image and an instruction, and gets back an action. Swap models without touching robot code.
- Serve is built for the real world: health checks, hot model loading, 4-bit quantization for tighter VRAM budgets, and first-class Prometheus metrics so you can see latency and throughput at a glance.
Features
- Model-agnostic backend — OpenVLA, SmolVLA, ACT, diffusion, or a custom class — selected by config, not code.
- Clean predict API — POST an image + instruction, get back an action vector. Your robot doesn't need to know which model is behind it.
- Hot loading & health — /health, /load_model and /predict endpoints let you swap or reload models with zero downtime.
- Fits your hardware — Optional 4-bit quantization runs large VLA models on modest GPUs. Bring your own — we help you size it, or recommend hardware that fits your budget.
- Observability first — Latency, throughput and GPU metrics export to Prometheus and the OhhO Fleet dashboards.
How it works
- Choose a model — Set the model class and checkpoint via environment variables.
- Launch the server — One command starts the FastAPI inference server.
- Call /predict — Your robot sends camera + instruction and receives an action.
- Watch it scale — Metrics stream to Prometheus / Grafana; raise call limits as you grow.
Specs
| Spec | Value |
|---|---|
| API | REST — /health, /load_model, /predict |
| Backends | OpenVLA, SmolVLA, ACT, diffusion, custom |
| Quantization | Optional 4-bit |
| Hardware | Your GPU (desktop, server or cloud) — we help you size it |
| Metrics | Prometheus /metrics endpoint |
| Deploy | Docker, single command |
Plans
| Plan | What this product gives you | Included |
|---|---|---|
| Spark | Not included | ❌ |
| Builder | 500 API calls / day | ✅ |
| Fleet | 10K API calls / day | ✅ |
| Forge | On-prem license, unlimited | ✅ |
Recommended plan: Fleet. Builder's 500 calls/day is for prototyping. If real robots are calling the model in production, choose Fleet for 10K/day — or Forge for an on-prem license with no metering.
FAQ
Can I run my own fine-tuned model? Yes. Serve loads any compatible checkpoint, including models you've fine-tuned with OhhO Data and the OmniVLA engine.
Where does inference run? On your GPU — desktop, server or cloud. Serve is software; you keep the model and the data.