Luma AI drops Ray 2 video model, teams up with AWS Bedrock

Two things usually ruin the fun of AI video: clips that cut off after three seconds, and rendering times that make you want to close the tab.

At AWS re:Invent, Luma AI announced its new flagship video generation model, Ray 2, while locking in a major cloud partnership with Amazon Web Services.

The update gives Luma's 25 million Dream Machine users longer generations, significantly faster renders, and direct access through enterprise infrastructure.

From 5 seconds to a full minute

The headline change in Ray 2 is duration. While the original model capped outputs at five seconds, Ray 2 generates 5-to-10 second cinematic clips and can stretch sequences up to a full minute.

Speed got a massive boost as well. The company claims the model can turn text and image prompts into finished video in as little as 10 seconds.

Under the hood, Ray 2 runs on a multimodal transformer architecture trained directly on raw video data. Luma says this approach improves physical intuition, allowing characters, animals, and objects to interact without melting into each other.

Camera controls also look much sharper. The model supports dramatic cinematic moves ranging from wide sweeping panoramas to tight close-ups, with less of the floaty distortion common in earlier visual models.

Moving into Bedrock and tapping custom silicon

Beyond consumer tools, Luma is chasing professional developers by bringing Ray 2 and its broader visual model suite to Amazon Bedrock.

Bedrock users can integrate Luma's video capabilities via a single managed API, skipping the headaches of hosting large multimodal models internally.

In return, AWS becomes Luma's preferred compute partner. Luma is training its foundation models on Amazon SageMaker HyperPod infrastructure to keep large-scale clusters stable.

Crucially, Luma is also porting workloads to Amazon's custom chips, AWS Trainium and Inferentia. The goal is to slash the sky-high inference and training costs associated with generative video.

Why it matters

Video models are notorious compute incinerators, and Luma's strategy looks like a masterclass in survival.

For creators, a 60-second limit paired with consistent characters moves AI video out of the meme generator category and into actual pre-production and storyboarding territory. A 10-second turnaround time makes iterating on prompts actually practical.

For the business, Luma has raised $80 million with a tiny engineering team. Going toe-to-toe against tech giants requires endless compute that venture dollars alone cannot sustain.

By anchoring to AWS and optimizing for Trainium silicon, Luma cuts its hardware overhead while gaining enterprise distribution through Bedrock overnight. It also hands AWS a credible high-end video model to counter rival cloud ecosystems.