Artificial Analysis (@ArtificialAnlys) on X

What it claims

Artificial Analysis reports that Hyperflow H3, Video Rebirth's 8-step LoRA for MiniMax H3 (released September 2026 with open weights), now supports Image to Video and Video Editing. It places #5 on AA-Video-I2V v2.1 and #3 on AA-Video-Editing v1.1. Its strengths sit in specific categories rather than evenly across the boards.

Evidence

In Image to Video, the thread scores two kinds of category. Capabilities are the model skills a prompt is written to test. Use cases are the real-world work the video is for.

• Multi-Scene & Narrative (capability): only Utopai X scores higher.
• Spatio-Temporal Consistency (capability): this is where Hyperflow H3 improves most over its base, MiniMax H3 (768p).
• Live-Action Film (use case): it ranks #2, behind only Dreamina Seedance 2.5.
• Marketing & Advertising (use case): it sits close behind the leader. The thread does not name the leader.

In AA-Video-Editing v1.1, each prompt pairs a source video with an edit instruction, and capabilities are edit types. Hyperflow H3 is strongest on Visual Effects Editing, where only Wan 3.0 scores higher. Overall it ranks above MiniMax H3 (768p).

How to use it

Hyperflow H3 is available on Video Rebirth's API at 1088p. It costs $9.00 per minute for Image to Video and $18.00 per minute for Video Editing. Dreamina Seedance 2.5 costs $34.12 and $40.82 per minute on the same two tasks. The weights are open, but the thread names no licence, repository or model size.

Limits

The fetched text is truncated at several "Show more" breaks. The lists of categories where Hyperflow H3 gains most over MiniMax H3 are cut off, for Image to Video use cases and for Video Editing alike. The thread gives ranks only. It gives no Elo or score values, so the gap to the leaders cannot be measured.

The base-model comparison is against MiniMax H3 at 768p, while Hyperflow H3 is priced and served at 1088p. Part of the gain may therefore come from resolution. The thread does not separate the two effects.

The page also does not cover:

• generation speed or latency,
• how the 8-step LoRA was trained,
• how self-hosting the open weights compares with the API.