AI is becoming continuous. Infrastructure should too.

Open Interaction builds inference infrastructure for realtime multimodal AI: video, voice and world models that stream in and stream out, in milliseconds.

Get started Talk to us

SCROLL TO OPEN

P50 FIRST FRAME 118 MSP99 164 MSACTIVE SESSIONS 2,418FRAMES / SEC 58,032GPU UTILIZATION 87 %TOKENS / SEC 1,204,000REGIONS 6

Request–response is over.

Interactive AI is realtime and long-running. Inputs keep arriving, outputs keep flowing, and a single session can stay open for minutes, hours, or indefinitely. The infrastructure has to work the same way.

Four families. One runtime.

Fast inference for the models that see, hear, generate and act, from video generation to world models.

$ oi models list --realtime

Video

Text and image to video, moving toward interactive video.

ID: video-realtime
MODELS: 6
FIRST FRAME: 120 ms
OUTPUT: 720p · 24 fps

READY

Voice

Realtime, full-duplex speech and audio.

ID: voice-duplex
MODELS: 4
FIRST FRAME: 180 ms
OUTPUT: 24 kHz · full-duplex

READY

Multimodal

Models that see, hear, generate and interact.

ID: multimodal
MODELS: 5
FIRST FRAME: 140 ms
OUTPUT: streaming tokens

READY

World models

Prediction for continuously changing environments.

ID: world-model
MODELS: 2
FIRST FRAME: 90 ms
OUTPUT: +1.2 s horizon

PREVIEW

Optimize the whole stack.

Model

Make the model cheaper to run without changing what it does.

Post-training
Distillation
Quantization
Model-specific kernels

Runtime

Serve long-lived streams instead of one-off requests.

Streaming execution
Continuous batching
Memory management
Stateful sessions

Hardware

Get the most out of every GPU in the fleet.

Kernel optimization
Hardware-aware execution
GPU utilization
Efficiency at scale

Faster inference. Same model.

Performance characteristics per stream on 1× H100 at 720p, 24 fps, against an unoptimized baseline. Dotted figures are placeholders until benchmarks are published.

Ref	Parameter	Typical	Conditions	Baseline
1.1	First-frame latency	120ms	p50 · video-realtime · 720p	2,100 ms
1.2	Tail latency	164ms	p99 · 60 min session	3,400 ms
1.3	Throughput	3.2×	concurrent streams / GPU	1.0×
1.4	Cost per output	−58%	GPU-hours / 1k frames	100%
1.5	Output quality	99.4%	benchmark score retained	100%
1.6	Session length	∞	stateful · resumable	one request

Baseline: the same model served request–response without optimization.OI-DS-001 · §04

From install to first frame.

Three steps. The runtime is open source, and hosted endpoints use the same API.

STEP 1

Install

pip install openinteraction

STEP 2

Authenticate

export OI_API_KEY="sk-…"

STEP 3

Open a stream

session = Client().stream(
  model="video-realtime")
# first frame in ~120 ms

Documentation DOCS
GitHub 2.4K STARS
Benchmarks REPRODUCIBLE

Build realtime AI on infrastructure made for it.

Get started Talk to us