Moondream
10kTry the model
NewPhoton 2.1: speech recognition, up to 3.1× faster

The world's fastest multimodal inference engine.

Up to 3x faster than vLLM and SGLang at every batch size from 1 to 8.
Powered by compiled megakernels. Free.

Benchmarks
$pip install moondream
Book a call
Models
Hardware
Quickstart
Sample code.
python
import moondream as md
from PIL import Image

model = md.photon("moondream3-preview")
answer = model.query(Image.open("aisle.jpg"), "Is the pallet blocking the aisle?")["answer"]
print(answer)

API reference · docs.moondream.ai/running-locally

Backed by
Felicis logoFelicisM12, GitHub-Microsoft Fund logoM12, GitHub-Microsoft FundAscend logoAscend
FAQ
What is Photon?

Moondream's inference engine for vision, audio, and language models. Each model is compiled into a single GPU program for its target chip. That compilation is where the speed comes from.

What's different than vLLM or SGLang?

vLLM and SGLang are built to maximize throughput over large request queues. Photon is built for realtime work: small batch counts, low latency, and steady tail latencies for systems that act on the next frame or the next utterance.

Is it free?

Yes. pip install moondream and run any supported model on your own hardware at no cost.

Which models and hardware are supported?

The support grid lists every model and chip Photon runs, and the validation status of each pair.

How are the benchmark numbers produced?

Every number on this site is generated from a published dataset: matched runs against vLLM and SGLang with pinned model revisions, raw CSVs, and the commands to reproduce them. See the benchmarks page.

Can Photon run my production workload?

That's what it's built for: realtime work like live video, robotics, and voice agents. Book a call and we'll go through your models, chips, and latency targets.