Luma Ray3

Ray3 is a video generation model that Luma AI released on 18 September 2025. Luma presented it as the world's first reasoning video model^[1]. The claim is about the process rather than about the pixels: according to Luma, Ray3 plans a scene and then checks and refines its own output, working with both text tokens and visual tokens^[1].

What "reasoning" means here

In language models, reasoning usually means spending extra computation on intermediate steps before producing an answer — drafting something, inspecting it, correcting it. Ray3's description applies the same shape to video: the model first lays out a plan for a scene, then looks at what it has generated and revises it^[1]. The useful part is the self-check. A model that only reacts to the prompt frame by frame has no way to notice that a shot has drifted; one that inspects its own intermediate result has something to correct against.

Picture quality and formats

Ray3 outputs HDR video in 10-, 12- and 16-bit ACES2065-1 EXR formats, and it can also convert existing SDR video into HDR^[1]. EXR is a file format common in film and visual-effects work, and the bit depths listed describe how finely brightness values are recorded — more bits leave more room for grading after the fact.

The model generates natively at 1080p and uses a neural upscaler to reach 4K; Luma states that the upscaler does not introduce blur or motion artifacts^[1]. At launch, native 1080p was available to select partners in early access^[1]. A Draft Mode produces quick test videos, so a shot can be iterated on cheaply before a final render^[1].

Where it launched

Ray3 shipped inside Luma's own Dream Machine and inside Adobe Firefly, in the Video module and in Boards^[1]. Adobe was the first partner to launch Ray3 outside Dream Machine, and it offered unlimited Ray3 generations for the first 14 days after launch^[1]. Other launch partners included Dentsu Digital in Japan and HUMAIN Create in the Middle East^[1]. Launching through a partner's interface as well as one's own product is a way to reach people who already work inside those tools.

Evaluation

On 14 October 2025, Luma published a Ray3 evaluation report^[4].

Later versions

Ray3.14 arrived on 26 January 2026. It brought native 1080p into core workflows, generation that is 4x faster and 3x cheaper than Ray3 at 720p, and clips of up to 18 seconds in Modify Video^[2].

Ray3.2, released on 9 June 2026, added up to 16 keyframes per clip, performance tracking for up to eight faces, 1080p clips of up to 20 seconds, and a full API^[3]. Keyframes give a creator anchor points to steer a shot; face tracking addresses the problem that faces are among the first things viewers notice going wrong; an API lets other software call the model directly.

Where it sits

Three things stand out about Ray3. The first is the reasoning framing: rather than competing only on resolution or clip length, Luma described the model in terms of how it plans and self-corrects^[1]. The second is production formatting — HDR in EXR with SDR conversion, native 1080p and a 4K upscaler — which speaks to finishing work rather than only to quick social clips^[1]. The third is distribution through Adobe Firefly alongside Dream Machine, which put the model in front of users of a widely used creative suite from day one^[1].

The version history points in a consistent direction: faster and cheaper generation, longer clips, more keyframes, face tracking and an API^[2][3]. Each of those is a practical constraint that shows up when generated video moves from demonstration to regular use.