Photon 2.0 targets physical AI inference and claims 2.3x throughput over vLLM

MikeBirdTech · x · 2026-08-04

Photon 2.0 targets physical AI inference workloads

Moondream says Photon 2.0 is a new inference engine built for physical AI, supporting Moondream, Qwen, and Gemma today, with more models coming.

What it claims

Photon 2.0 is designed for workloads like robots, cameras, industrial systems, and computer-use agents.
The team says these workloads differ from chat: they have low concurrency, stricter latency needs, and often serve live perception.
In matched benchmarks on NVIDIA H100, Photon reportedly beats vLLM and SGLang on every throughput test.
The company says it can reach up to 2.3× higher throughput and bring models online faster.

Why the company says a new engine is needed

Existing inference stacks are optimized for batching prompts and maximizing tokens per second across GPU fleets.
Physical AI, by contrast, continuously ingests images, audio, and video and often sacrifices throughput for response time.
Moondream argues the combination of model × chip × deployment objective is too varied for hand-tuned kernels to scale well.

Bigger picture

Photon 2.0 is positioned as part of the infrastructure stack for deploying perception-heavy AI systems across edge, on-prem, and cloud environments.

More from Infra