AI&Design
All archives
AUGUST 2024 — NOVEMBER 2026 · KST

MiniMax Hailuo

Compare Video-01, Hailuo 02, Hailuo 2.3, and MiniMax H3 across MiniMax, Runway, and Higgsfield. Follow the differences between a base model, its open weights, and a post-trained fast variant.

Posts archived
Attachments examined
Capabilities with evidence

What changed, when

Announcement dates in KST · Select a change or open its example.

Result example attached Announcement or contextual mediaBlank lane · no qualifying announcement in this source

Workflow threads

Read the dated updates and sources
  1. · MiniMax releases Video-01

    Generate a clip from text or combine a starting image and description to guide the scene. MiniMax's article is dated August 31. Its 2025 Hailuo 02 release describes the early demo reaching creators in late August.

  2. · Hailuo 02 adds native 1080p generation

    MiniMax improves instruction following and demonstrates difficult physical motion, such as gymnastics, alongside native 1080p output. The release film is assembled from several 6–10 second model clips.

  3. · Hailuo 2.3 improves motion, expression, and style

    The model demonstrates physical motion with camera movement, more distinct visual styles, and character micro-expressions. A Fast variant targets high-volume image-to-video generation. The model notice also introduces a Hailuo Media Agent update. It is a product workflow distinct from model output.

  4. · H3 adds mixed references and native stereo

    Use text, images, video, and audio as a single context. H3 generates up to 15 seconds, supports 2K, and produces a synchronized stereo soundtrack. The research release describes multiple execution routes and capabilities. Individual product endpoints support different subsets.

  5. · MiniMax publishes H3 open weights

    The release covers text-to-video with audio, first and last frames, and multimodal reference generation. A separate 2K regeneration route is listed. MiniMax keeps its hosted multi-stage Context IR workflow outside the open release.

  6. · Runway Dev adds Hailuo 3.0

    The API supports generation from text, image or video, with keyframes and image, video, and audio reference inputs. Runway documents 5–15 seconds at 768p or 2K. The API ID is hailuo3.

  7. · Higgsfield documents H3 generation modes

    Use the model from text, first and last frames, or a collection of reference assets. Higgsfield documents available sizes and generated audio. The article's structured publication timestamp converts to August 17 in Korea. The article is evidence of availability by publication.

  8. · fal introduces a faster H3 post-train

    fal post-trains MiniMax H3 for prompt adherence and lower latency. The showcase covers text, first and last frames, and reference generation. Speed and ranking figures describe fal's test setup. Each integration exposes a subset of the variant's modes.

  9. · Higgsfield adds H3 Max frame transitions

    Generate from a prompt or source image and guide transitions with a start and end frame. The image-to-video route aims to preserve composition, identity, and aspect ratio. Higgsfield documents a 768p cap for Max. Its listing distinguishes Max from the base H3.

  10. · Runway Dev adds H3 Max with prompt expansion

    Generate 5–15 seconds from a prompt or start image, with an optional last frame. Prompt expansion ranges from disabled to a quality rewrite. This variant uses h3_max. API resolutions are 480p and 768p.

  11. · Recraft adds MiniMax H3 Highspeed and Lite

    Both video variants generate 4–15-second clips with sound at 480p or 768p. Lite is the lower-cost option for draft work. Video and external-model access is required on Recraft Studio. H3 Image variants in the same notice are separate from these video models.

Documentation and availability notes
  • MiniMax API · The current API documents reference, video, and audio inputs and separates models by endpoint capability.
  • MiniMax open weights · The model card indexes the released bases, supported conditioning modes, and 2K regeneration.
Coverage through November 20, 2026. Exact publisher timestamps are shown in KST; date-only notices retain the publisher's day.Back to capability timeline