Aggregate disclosure leaves the sources opaque
Black Forest Labs says FLUX 3 jointly learns from images, video and audio and can generate video with audio up to 20 seconds. Its 77% preference rate against Runway Gen-4.5 and 93% against Luma Ray 3.2 come from the company's early-access evaluation; the harness remains in development and no independent replication is published.[1]
The company does publish aggregate information about its training data. Its June 9 notice describes a mix of text and image data for the FLUX family, including third-party acquisitions, contractor-created, usage, synthetic and internally generated data. The FLUX-mimic post also reports tens of millions of hours of general video and hundreds of thousands of hours of manipulation video. What remains missing is a traceable, dataset-level inventory of FLUX 3's video and audio sources, including licensing, consent and removal terms.[3], [2]
The distance between allegation and evidence
White House science adviser Michael Kratsios alleged that Moonshot developed Kimi K3 by copying Anthropic's Fable, but provided TechCrunch no evidence or technical detail. The roughly two weeks between Fable's July 1 availability and K3's release are, according to researchers interviewed, too short for comprehensive distillation from Fable. That skepticism weakens the allegation; it does not prove Moonshot used no earlier Anthropic output.[4]
For the allegation to be persuasive, the public would need access and payment logs, timestamped prompt-response matches, the attribution method's error rate and a training chronology. None of that has been disclosed. The researchers' timing objection weakens the accusation, but it does not explain how K3 was trained.[4]