Ray3 by Luma AI: status and timeline

Milestones

No announced next step

Ray3 development begins2024Complete.
Ray3 released with reasoning capabilitiesSep 2025Complete.
Series C funding closesNov 2025Complete.
Ray3.14 released, 4x faster with lower costsJan 2026Complete.
Adobe Firefly integrationSep 2025Complete.

Most important updates

Current obstacles

Long-form video processingCurrent models efficiently process clips of seconds to minutes. Feature-length video understanding remains computationally challenging.
Inference latencyReal-time video reasoning for live applications requires dramatic reductions in processing time per frame.

Physics limits

Memory demands scale with video lengthStoring and processing longer videos requires more GPU memory. This limits practical video duration or forces chunking strategies that lose context.
Training data diversity affects generalizationVideo models are only as good as their training data. Rare scenarios, specialized domains or novel combinations require either retraining or fine-tuning.

How it works

3 parts

Video encoder

Processing temporal sequences

Unlike image models that see single frames, Ray3 encodes entire videos as sequences, learning how scenes evolve over time.

Reasoning layer

Causal inference

The model learns to identify cause-and-effect relationships in video: what action led to what outcome, which enables prediction and planning.

Multimodal fusion

Combining vision, audio and text

Ray3 integrates information from video (visual content), audio (sounds, speech) and text (captions, metadata) for holistic understanding.

Spec sheet

Spec	Ray3
Series C round	$900M^R (reported)
Dedicated supercluster	2 GW^R (reported)

^R reported by the company

Update log

3 updates

About Luma AI

The team behind Ray3

Luma AI

Ray3 video reasoning, generative video, multimodal AI

Luma AI develops multimodal foundation models for video generation and reasoning, with Dream Machine and Ray3 as flagship products. The company raised $900M in Series C in November 2025 to train models on a 2 GW Saudi supercluster for advanced video understanding.

Ray3 released September 2025 as a video reasoning model. Ray3.14 released January 2026 for faster generation at lower cost.
$900M Series C in November 2025 led by HUMAIN supports training on 2 GW Saudi supercluster for AI.
Builds multimodal foundation models for video generation, understanding, reasoning for content creation and automation.

lumalabs.ai (opens lumalabs.ai)Company profile →

Founded: 20215 yrs
Headquarters: United States
Status: Private
Valuation: $4BprivateNov 2025
Raised: $963M3 rounds
Last round: Series C · $900MNov 2025
Works in: AIImage & voice generation
Coverage: 1 program · 3 updateslatest 1 Jan 2026checked 10 Oct
Lead investors: HUMAINAndreessen HorowitzAmplify Partners
Partners: HUMAINAdobe
People: Amit JainCo-founder and CEO

Spec

Ray3

Series C round

$900M^R (reported)

Dedicated supercluster

2 GW^R (reported)

Ray3

Milestones

Most important updates

Current obstacles

Physics limits

How it works

Processing temporal sequences

Causal inference

Combining vision, audio and text

Spec sheet

Update log

Jan 2026

Nov 2025

Sep 2025

About Luma AI

Current obstacles

Physics limits

How it works

Processing temporal sequences

Causal inference

Combining vision, audio and text

Spec sheet

Update log

Jan 2026

Nov 2025

Sep 2025

About Luma AI