The open video model tops out at fifteen seconds, not twenty-six

MiniMax's H3 video model has drawn 2,900 stars and 171 forks in the four days since its weights went public, and the enthusiasm has outrun the specification. The company's own repository caps a single generation at fifteen seconds at a 768-pixel short side, says 2K output requires a second pass through a separate stage, and states that the component most responsible for output quality is hosted and deliberately excluded from the release.

Key facts

• Output ceiling: 4 to 15 seconds per generation, default short side 768 pixels, 32 kHz stereo audio. 2K requires the separate H3-Regenerate-2K stage.
• What shipped: two checkpoints - FL2VA (first-and-last-frame) and Ref2VA (omni-reference) - each with its own processor, tokenizer, text encoder, and visual and audio autoencoders.
• What did not: H3-Context-IR, which MiniMax describes as "a hosted preprocessing and orchestration system" and says is "critical to the quality of the final output."
• Primary source: the MiniMax-H3 repository, initial commit August 5, 2026, still receiving fixes as of today.

H3 is a genuinely ambitious system and it is worth being precise about what it does. MiniMax calls it "a general-purpose, omni-modal generative system" that takes a mixture of text, images, video, and audio as context and generates video with native stereo sound. That last part is not common: most open video models produce silent clips that someone else scores afterwards. H3 generates the audio and the picture together, at 32 kHz stereo, which is why the reference mode accepts audio clips as inputs alongside images and video.

The architecture has three stages, and understanding the split explains the whole story. H3-Context-IR reads whatever messy pile of references and instructions a user provides and works out what they mean - which image is the character, which video is the motion reference, what the audio is supposed to be doing, how it all relates to the intended shot. It compresses that understanding into a structured Context Intermediate Representation. H3-Base takes that representation and generates 768p video and audio from it. H3-Regenerate-2K feeds the 768p result plus the original context back through the model to produce a higher-resolution version with better detail.

Only the middle stage came out. MiniMax's stated reason is technical rather than coy - "H3-Context-IR relies on a multi-stage workflow and multiple hosted models and services" - and it offers an API to reproduce that behavior plus a prompt-writing guide for people who want to build their own. But the company does not soften what that means for output quality: it "strongly recommend[s] incorporating it into your generation pipeline."

The analogy is a film crew. What is now downloadable is the camera and the sound rig, and they are real, professional equipment. What stayed behind is the director and the script supervisor - the part that reads the brief, decides what the shot is supposed to be, and hands the crew something coherent to execute. You can absolutely shoot without them. You will notice.

None of this makes the release small. Two full checkpoints with their own encoders and decoders, deployment paths for the major inference servers, and a permissive-enough community license is a serious drop, and the fork count says builders are wiring it up rather than just admiring it. The interesting question was never whether the weights were real. It was whether the gap we flagged when the weights first landed had closed in the week since. On MiniMax's own documentation, it has not: the hosted stage is still hosted, and the duration ceiling is a property of the model rather than of anyone's graphics card.

That ceiling is where the community framing and the specification part company. Fifteen seconds is the documented maximum for one generation. Longer results are not a vanilla single-pass output - they are stitched, regenerated across passes, or produced by a wrapper doing extra work around the model, and any of those is a legitimate thing to build but is not the same claim. The same applies to resolution: 2K is a second stage, not a setting.

The honest caveat cuts in the model's favor as well as against it. MiniMax says the initial release provides full-attention inference only, with sparse attention deferred to a later update - which means the version people are benchmarking on their own hardware today is the expensive one, and the efficiency work that makes long contexts practical has not shipped yet. Local runtimes have a track record of closing exactly this kind of gap; llama.cpp merged MiniMax's sparse attention for a previous model after a similar wait. Judge the hardware requirements again when that lands. Until then, the accurate summary is that a capable open-weight video-and-audio generator now runs outside MiniMax's data center, at up to fifteen seconds a shot, without the part that makes the hosted version look as good as it does.

---

Originally published on Ground Truth, where every claim is checked against the primary source.