
Reactor Serve
Serverless infrastructurefor real-time AI.
Deploy, optimize and build with real-time models.
Deploy real-time AI. Anywhere.
Bring any model, deploy it to GPUs reserved for you, and run real-time AI from a browser, a phone or a robot.
from reactor_runtime import ReactorApp, MediaInput, Output
from reactor_runtime import Video, event
class Inputs(MediaInput):
webcam: Video
class Outputs(Output):
main_video: Video
class MyWorldModel(ReactorApp):
media: Inputs
fps = 30
def load(self, config_path):
self.pipe = load_pipeline(config_path)
def generate(self, frame):
return self.pipe.step(frame, self.prompt)
@event(name="set_prompt")
def set_prompt(self, prompt: str):
self.prompt = prompt- Reactor runtimeRuns and optimizes your model
- Your inference code
- Your weights
$ reactor model deployUsing model spec from reactor.yaml (world-explorer:v1.3.0)Applying model configuration... ✓Deploying release v1.3.0 to us-west, us-east, eu-west... ✓ ✓ Deployed Model world-explorer Release v1.3.0 Rollout rolling, no sessions dropped
import { Reactor } from "@reactor/sdk";
const reactor = new Reactor({
modelName: "world",
jwt,
});
await reactor.connect();
reactor.publishTrack(gamepad);
reactor.sendCommand("set_prompt", {
prompt: "night falls",
});Bring your own stack
Diffusers, custom CUDA kernels or a VLA policy: if it runs in Python on a GPU, it runs on Serve.
Any modality
Stream out video, audio, joint targets or JSON at the real-time rate you set.
Your code, your rules
State, moderation and command handling stay plain Python you own, with no framework to learn.
If it’s real time, it runs on Reactor.
Build the next generation of applications with models that respond as people, environments and interfaces change.

World models
Interactive worlds for games, simulation and training data, generated in real time.
Avatars
Video agents for support, tutoring and companionship that listen and reply face to face.

Real-time video
Generate video and audio faster than real time for live streams, ads and editing tools.

Robotics
Run robot policies and world models in the cloud, connected to your robots in real time.
Explore Reactor Robotics
+And more
Have another use case in mind? Bring your model and run it on Reactor.
Tell us what you’re buildingClose to your users. Built for low latency.
Global by default
An extensive network of compute regions worldwide. Deploy close to your users across major clouds.
Production-grade streaming
Stream video, audio and data in both directions over WebRTC, QUIC or gRPC. Send inputs and commands to your model and receive outputs as they’re generated.
Performance at every scale.
Scale across regions
Add GPU capacity across regions, right from one platform. Your model code stays the same.
Get more from every GPU
Our inference team tunes your model with custom kernels and a real-time engine to serve more sessions per GPU.
Instant deployment
Deploy your world model with one command. We take care of GPU capacity, real-time connections and monitoring.
$reactor model deploy✓ Configuration validated✓ GPU capacity provisioned✓ Model release deployed
Deployment ready.
Waiting for connections...
The right compute for your real-time model.
Choose NVIDIA GPUs or AWS Trainium for your model.
NVIDIA Blackwell
For the largest world models and video pipelines at full frame rate.
- B200180 GB HBM3e
- B300270 GB HBM3eOn request
- GB200186 GB HBM3eOn request
- GB300279 GB HBM3eOn request
- RTX 6000 Pro96 GB GDDR7
NVIDIA Hopper
Where most real-time models land, with one H100 per instance by default.
- H10080 GB HBM3Default
- H200141 GB HBM3e
AWS Trainium
For models compiled for Neuron, in AWS regions.
- Trainium 3144 GB HBM3eOn request
More hardware coming soon.
Reserved
Capacity committed to your model. Choose your hardware and region, and keep it ready for steady traffic.
Reserve capacityOn demand
Capacity available at a moment’s notice. Pay by the minute for the compute your sessions use.
Get startedPowering real-time world model sessions for
Reactor Serve
The next generation of AI runs in real time.
Bring your world models, avatars, and robot policies to production with inference, GPU scaling, and streaming on one platform.
[email protected] ↗





