Skip to content
NewGR00T N1.7 runs 24.6% faster on H100 with Fenon.

Deploy Humanoids like Software.

Cut deployment time. Share compute across your fleet. Keep safety on every robot. One API for frontier models.

GPU H100 GPU MI300X GPU B300 ARM

From perception to action.

Turn images and video into structured understanding, or robot observations into the next action. One API for both.

Perception models

Extract hand pose, object detections, depth, segmentation and more from your images and video. Run across your fleet’s recordings and write structured results to your storage.

Clips analysed
3,008 / 4,812
In parallel
12
Models per clip
4

Action models

Run policies and vision-language-action models on your robot’s observations. Get the next actions back, with every action checked on the robot first.

Actions returned
0
Run by robot
0
Safety stops
0
Fenon vs robot rate
1.5×
  • 72Pending
  • 0Generated
  • 0Executed
  • 0Blocked

Different models.
One inference layer.

Policies, world models and perception, served through the same API, live or in batch.

Policies & VLAs

Send what the robot sees, and an instruction if your model takes one. Get the next actions back.

World models

Give it the current scene and an action. Get back what happens next, frame by frame.

Perception

Detection, segmentation, depth and pose from your images and video.

Your policy loop.
Powered by Fenon.

Bring your models to Fenon. Go from simulation to inference with one Python client.

uv pip install fenon
robot_loop.py
from fenon import Fenon client = Fenon(    api_key="fn-sk-***",)obs, info = env.reset() while True:    result = client.run(        model="nvidia/groot-n1.7",        input=obs, mode="live",    )    action = result.action     obs, _, done, truncated, _ = env.step(action)    if done or truncated:        obs, info = env.reset()

Built for physical AI.

Latency-sensitive requests, multimodal inputs and heavy video processing change the infrastructure problem.

  • Architecture

    The control boundary, the transport, and what stays on the machine.

  • Latency

    What a 56 ms round trip is spent on, and where it can be recovered.

  • Model choice

    Physical Intelligence, NVIDIA, Dreamlab or your own model. Run your choice through one API.

Tell us what you’re building and which models you need. Fenon handles the inference layer.