Title card reading 'WeMM-Video-9B vs HiDream-O1-Video-1.0' with the subtitle 'Two narrow waists', two panels split by a vs disc: the left headed WeMM-Video-9B with an 'open weights' badge and rows reading 'no Elo, no board', 'reads video, returns text', '19.54 GiB BF16', 'Apache 2.0', 'unannounced'; the right headed HiDream-O1-Video-1.0 with a 'metered API' badge and rows reading 'Elo 1,101, rank 14', '6,976 votes', '$5.80 per minute', 'one reference image', '10 s maximum'.
Guides & Insights

WeMM-Video-9B vs HiDream-O1-Video-1.0: Two Narrow Waists

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

Both of these models are narrower in production than they are on paper, and they are narrowed by different things. HiDream-O1-Video-1.0 is narrowed by its own API: the announcement describes text, image and video input with five-to-twenty-second 1080p output and natively synchronised audio, and the HiHarness endpoint that actually exists accepts exactly one reference image, an optional prompt, and a boolean. WeMM-Video-9B is narrowed by having no serving contract at all — Tencent's WeChat Vision group published it on 17 November 2026 as Apache-2.0 weights and an arXiv preprint, 2611.26409, with no endpoint, no price and no announcement, so the sixteen-frame limit in its shipped vLLM recipe is not a policy anyone can appeal, it is just what the repository does when you follow its instructions.

The comparison that follows is therefore not about capability. It is about which constraint you can design around, and the answer turns out to depend on whether you would rather pay for a documented ceiling or host an undocumented one.

What the paper says and what the API accepts

HiDream-O1-Video-1.0 released on 28 August 2026 and reached the image-to-video board on 10 September. The vendor's announcement is broad: text, image and video input, five to twenty seconds of 1080p output, natively synchronised audio, narrative-planned duration, physics constraints.

The HiHarness API is not broad. It takes one reference image — a URL or Base64, no larger than 20 MB and no smaller than 40 KB — plus an optional text prompt, plus a force_10s boolean. There is no resolution selector. There is no reference-video field. There is no audio control. If you want five seconds you get ten; if you want twenty you get ten. The gap between the paper and the endpoint is the most concrete fact about this model and the one most likely to change your architecture.

• Input — HiDream-O1-Video-1.0: one reference image plus optional prompt. WeMM-Video-9B: video, audio, images and text, up to four images, two videos and two audio clips per request.

• Output — HiDream-O1-Video-1.0: 1080p video with synchronised audio, ten seconds in practice. WeMM-Video-9B: text in English or Chinese.

• Constraint — HiDream-O1-Video-1.0: the API surface. WeMM-Video-9B: the shipped sampling recipe, 16 frames at 2 fps.

Two-column scoreboard headed 'WeMM-Video-9B vs HiDream-O1-Video-1.0 — the scoreboard', six shared dimensions in rows: Price ('$0.00, no endpoint' against '$5.80 per minute'), Board position ('none, unrankable' against 'Elo 1,101, I2V rank 14'), Input ('video, audio, images' against 'one reference image'), Output length ('text only' against '10 s in practice'), Audio ('reads the soundtrack' against 'synchronised in-pass'), Release ('17 Nov 2026' against '28 Aug 2026'), with a footer reading 'Artificial Analysis AA-Video-I2V v2.1 read 20 November 2026; WeMM-Video-9B figures vendor-reported.'

One board row, one empty cell

HiDream-O1-Video-1.0 has exactly one independent measurement, and it is a good one to argue with.

• AA-Video-I2V v2.1 — 14th at Elo 1,101 ±14 over 6,976 samples, August 2026 release, $5.80 per minute. That is a tight interval on a large sample, and it sits between Wan 3.0 at 1,109 and Agnes-Video-2.5 at 1,105 — a crowded band where fourteen points is one interval.

• Absent from AA-Video-T2V v2.1 and AA-Video-Editing v1.1. The text-to-video absence is the interesting one, because the vendor's own announcement leads with text input. A model with no text-to-video row and no reference-video field has a text-to-video story that exists only in prose.

• WeMM-Video-9B — no row on any of the three boards, for the structural reason: all three rank generated footage, and it generates none. An empty cell, not a low score.

At $5.80 a minute, HiDream-O1-Video-1.0 is one of the cheaper entries on the image-to-video board. Only Pika H3 Flex at $2.10, HeyGen Video 1 at $2.40, Agnes-Video-2.5 at $2.40, Vidu Q4 Preview at $5.04 and P-Video-2-Pro at $5.70 come in under it among the ranked field, and it beats several models priced three to four times higher.

Screenshot of the Artificial Analysis image-to-video leaderboard, AA-Video-I2V v2.1, with ranked rows carrying Elo, confidence interval, vote count, release month and per-minute price, including HiDream-O1-Video-1.0 at rank 14 on Elo 1,101 over 6,976 samples at $5.80 per minute, between Wan 3.0 at 1,109 and Agnes-Video-2.5 at 1,105.

Nineteen and a half gigabytes of detail

Where WeMM-Video-9B is generous is in what it tells you about itself. The Hugging Face record lists 10,488,603,384 parameters across 1,305 tensors — a 7.94B language backbone including a 1.02B token embedding, an untied 1.02B output head, a 1.08B audio tower of 32 Whisper-style layers at width 1,280 behind a two-layer Q-Former at ten audio tokens per second, and a 0.46B vision tower of 27 blocks at hidden 1,152 with patch size 16 and a 2×2 spatial merge.

That precision is a form of honesty the HiDream side does not offer: Black Forest Labs and HiDream both decline to publish parameter counts, and neither is obliged to. But it also sets a trap, because a detailed architecture invites the assumption that the shipped configuration matches it. It does not.

• Processor defaults — 2 fps, minimum 4 frames, maximum 128 frames, 100,663,296 total pixels.

• Shipped vLLM recipe — 2 fps, 16 frames, 262,144 pixels, sampled around the midpoint of the clip.

• Context — 8,192 tokens on every serving command, against 262,144 position embeddings in the architecture.

Sixteen frames at two per second is eight seconds of coverage, and the sampling begins in the middle. A sixty-second clip loses its first and last twenty-six seconds. The architecture would allow far more; the recipe does not, and the recipe is what ships.

Screenshot of the Hugging Face model page for tencent/WeMM-Video-9B showing the image-text-to-text pipeline tag, Transformers and Safetensors libraries, English and Chinese languages, the arXiv identifier 2611.26409, an Apache 2.0 licence, a model card overview stating 19.54 GiB of BF16 weight files, and a file panel reading 10B params and BF16 tensor type.

Ten seconds against eight

Put the two caps next to each other and the shape of the decision appears. HiDream-O1-Video-1.0 gives you ten seconds of finished 1080p with synchronised audio, no matter what you ask for. WeMM-Video-9B looks at eight seconds of footage you already have, no matter how long it is.

Both numbers are arbitrary. HiDream's ten seconds is a product decision embodied in a boolean; WeMM-Video-9B's eight is a default in a recipe file. Neither is a law of physics. But only one of them comes with a vendor who will tell you what happens at the boundary, and it is not the one distributed as a checkpoint.

If you generate a ten-second clip with HiDream-O1-Video-1.0 and then ask WeMM-Video-9B what happened in it, the describer samples eight of those ten seconds from the middle. Two seconds — the first and the last — never reach the model. That is a coordination problem you solve with a config change, and it is exactly the kind of coordination problem this pair creates and neither vendor documents.

Model fusion, and where it earns the sentence

Neither model is hosted at OrcaRouter. HiDream-O1-Video-1.0 is reachable through HiDream's own platform and several third-party platforms; WeMM-Video-9B is reachable by downloading it.

The reason a comparison like this one ends up in an orchestration conversation is that the two halves fail differently. The hosted half fails loudly — a rate limit, a timeout, an error code — and that is what automatic failover is for. The self-hosted half fails quietly: it returns a confident description of eight seconds of footage and says nothing about the twenty-six seconds it never saw. Model fusion is the more interesting tool here, because a second model reading the same clip under a different sampling policy will disagree at the boundaries, and the disagreement is the only signal you get that the window was too small.

For the hosted leg, provider list price passes through at 0% markup, which matters more than usual at $5.80 a minute: a vendor price change on a cheap model is a larger percentage swing than on an expensive one, and tracking the published rate rather than a cached copy keeps a hundred-clip batch forecast honest.

Which constraint you can live inside

Choose HiDream-O1-Video-1.0 if one reference image and ten seconds is the job. You get a documented endpoint, synchronised audio, a $5.80 rate and a board row at 14th on image-to-video. You give up text-to-video, resolution control and any duration other than ten seconds.

Choose WeMM-Video-9B if describing footage is the job. You get a 9B omni-modal describer with strong reported audio-visual numbers, an Apache-2.0 licence and a parameter count you can audit. You give up a serving contract, independent verification, and any guarantee that the eight-second window is wide enough for the clips you actually have.

What you should not do is compare the two on quality. Neither has been evaluated on the same instrument, and one of them cannot be. The comparison that pays is between two narrow waists — a boolean and a recipe file — and which one you can build around without lying to yourself about what it sees.