Title card reading 'WeMM-Video-4B vs HiDream-O1-Video-1.0' with the subtitle 'Two kinds of narrow', two panels split by a vs disc, the left subtitled 'open weights, no endpoint' and listing 5.61B parameters, Apache 2.0 and patched vLLM or SGLang, the right subtitled 'hosted, one input' and listing unpublished parameters, one reference image and $5.80 per minute, with a footer reading 'Artificial Analysis board read 20 November 2026.'
Guides & Insights

WeMM-Video-4B vs HiDream-O1-Video-1.0: Two Kinds of Narrow

Author

jinhao.song

Date Published

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

HiDream-O1-Video-1.0 and WeMM-Video-4B both describe themselves more broadly than they behave. HiDream's launch release advertises a model that takes text, images and video and returns five to twenty seconds of 1080p footage with synchronised audio; the documented API accepts exactly one reference image and an optional prompt. WeMM-Video-4B takes text, images, video and audio and returns text, which is accurate — and the model is not served anywhere, so what it accepts in theory is what you get only after patching your own inference engine. One is narrowed by its serving contract, the other by having no serving contract at all. That is the whole comparison, and everything below is detail hanging off it.

Neither model is one you can call today in the mode its paper emphasises. How much that matters depends on which half of the gap you can close yourself, and that is a question about licences and GPUs rather than about quality.

The row, and the version it sits on

HiDream-O1-Video-1.0 has an independent measurement. Read on 20 November 2026, on the AA-Video-I2V v2.1 board, it sits 14th at Elo 1,101 with a 95 percent interval of ±14, which is a range of 1,087 to 1,115, over 6,976 votes, recorded as an August 2026 release, at $5.80 per minute of generated video.

Six weeks earlier, on the v1.0 board, the same model was sixth at 1,175 ±10 over 3,231 votes. Read carelessly, that is a collapse. It is not, and the reason is worth stating plainly because it is the most common error made with these boards: v1.0 and v2.1 are different instruments with different Elo scales, different vote pools and a different taxonomy. A score does not transfer between them, so the distance between 1,175 and 1,101 is not a measurement of how much worse the model got. What did change is real — rank 14th is a worse place to be than rank 6th under the current board's rules — but the arithmetic that connects them does not exist.

WeMM-Video-4B has no row on any board, and the absence is structural rather than a pending review. The AI video boards rank generated footage. A model that reads video and writes sentences cannot be entered in a vote between two clips. Until an omni-modal video-understanding arena exists, every WeMM-Video-4B number is Tencent's own.

What HiDream's API actually accepts

The serving surface is documented and it is narrow, in ways that map directly onto the launch claims.

• Input — exactly one reference image, given as a public URL or Base64 between 40 KB and 20 MB. The prompt is optional and defaults to empty. The announcement lists video among the model's inputs; the request schema has no field for it.

• Resolution — no selector. The result preserves the reference image's aspect ratio. The 1080p claim in the announcement is not steerable from the request.

• Duration — one boolean, force_10s, defaulting to false. Left false the model chooses; set true you get ten seconds. The advertised five-to-twenty-second range has no numeric field.

• Audio — no audio input or audio-generation parameter is exposed, despite the native synchronised audio the release describes.

• Result contract — asynchronous. You submit, get a task_id, and poll. result.status = 1 means every subtask reached a terminal state, not that generation succeeded; success is sub_task_results[].task_status = 1, with 3 for generation failure and 4 for safety review.

So the model you can call is an image animator with two duration modes. The model described in the release is a joint text-video-audio generator with planned duration. Both statements are documented by the same company, and the difference is a roadmap rather than a contradiction.

What WeMM-Video-4B asks you to build

The other side has no contract to be narrow, because there is no endpoint. WeMM-Video-4B is 5,611,746,312 parameters — 10.45 GiB of BF16 weights — released under Apache 2.0 in a family of four sizes that all appeared inside a nineteen-minute window on 17 November 2026, with the technical report, arXiv 2611.26409, following on 19 November.

Getting it to answer takes work that HiDream's users never see. Neither vLLM nor SGLang supports the architecture upstream; the repository ships source-overlay patches for four exact upstream releases — vLLM 0.18.0 and 0.31.0, SGLang 0.5.16 and 0.5.19 — each pinned to a specific PyTorch, Transformers and CUDA 12.9 combination, with Python 3.12, FFmpeg, and a C++ toolchain plus Ninja for SGLang's first-use kernel compilation. The GGUF bundle is built for a native runtime you compile from the repository, not for llama.cpp, and Tencent labels it exported_unvalidated.

What you get for that is a describer with an unusually large audio stack: 4.21 billion parameters of Qwen3.5-4B-derived language model, 0.33 billion of vision tower, and 1.07 billion of Whisper-style audio encoder-decoder feeding a Q-Former at 10 tokens per second. Nearly a fifth of the checkpoint is ears, which is why its joint audio-visual scores are the interesting ones.

The modes each model is compared in

A fair contrast here is not "which is better" but "which mode is the one you can actually reach", and the two answers land in different places.

• HiDream-O1-Video-1.0 — reachable mode: image-to-video, one reference frame, fixed aspect, ten seconds or model's choice, no audio control. Scored mode: image-to-video. The board and the API agree.

• WeMM-Video-4B — reachable mode: video with soundtrack in, text out, at 16 sampled frames and 8,192 tokens as shipped. Scored mode: the paper reports Thinking mode on the video tables, Instruct on the audio table, and Thinking again on image and text, and it does not state the frame count or context length used.

• Both — vendor-reported where a number exists, and no third party has reproduced either. HiDream's advantage is that one configuration has been voted on by people who do not work for HiDream.

Two-column contrast headed 'WeMM-Video-4B vs HiDream-O1-Video-1.0', the left listing open weights under Apache 2.0, video and audio in with text out, a patched vLLM or SGLang overlay and no independent score, the right listing a hosted API, one reference image with ten seconds or the model's choice, vendor-hosted runtime and 14th at Elo 1,101, with a footer reading 'Artificial Analysis board read 20 November 2026; WeMM figures vendor-reported.'Screenshot of the Artificial Analysis image-to-video leaderboard, AA-Video-I2V v2.1, with HiDream-O1-Video-1.0 at rank 14 on Elo 1,101 with a plus-or-minus 14 interval over 6,976 votes, an August 2026 release and $5.80 per minute.

The audio question, which only one side can answer

HiDream generates audio and WeMM-Video-4B understands it. Those sound like the same capability in a feature list and they are not, and for a team building a pipeline the distinction decides which one is even relevant.

On the generation side, HiDream's release describes a single process that models text, video and audio together rather than adding sound afterwards, so lip movement, ambience and physical action share a timeline. Whether you can steer that is a separate matter: the documented request exposes no audio parameter, and the announcement does not publish a track count, so dialogue, music and effects as separable stems are unverified on both counts.

On the understanding side, WeMM-Video-4B scores 83.9 on Video-MME with audio against Qwen3-Omni-30B-A3B-Instruct's 77.3 and MiniCPM-o-4.5-9B's 69.2, and 82.5 on Daily-Omni against 70.7 and 80.2. It loses the pure-audio benchmarks to a proprietary flash model — MMAU 75.2 against 80.4, MMAR 67.3 against 74.0 — with one win on MMAU-Pro at 68.1 against 64.8. Those are Tencent's numbers from Table 2, run in Thinking mode, and nobody outside the lab has repeated them.

So the honest summary is that one model makes sound you cannot yet inspect and the other inspects sound you have to supply. Neither is the audio tool a production pipeline wants until the first gains an audio parameter and the second gains a host.

Where a router helps and where it cannot

OrcaRouter serves neither model. There is no WeMM endpoint — the catalogue holds no WeMM checkpoint — and no HiDream route either; HiDream-O1-Video-1.0 is reachable through HiDream's own platform and several third-party platforms. Nothing below is a claim that we host either one.

The place a router earns its keep in this matchup is the migration itself, and the tool for that is the routing DSL rather than the model list. Both of these models sit behind a serving surface that is visibly provisional — one is a single-input preview with a two-state terminal flag, the other is a source overlay pinned to specific upstream releases — and a stack that hard-codes either is a stack that breaks on somebody else's release note. A routing policy that names a primary and a fallback, and that can compare two models inside one request rather than two integrations, turns that risk into configuration. That runs over one OpenAI-compatible endpoint across more than two hundred models at provider list prices passed through with 0% markup, so a vendor price change moves the same day. For the vision-only part of the WeMM workload, Qwen3-VL-235B-A22B-Thinking and Qwen3-VL-8B-Thinking are both on the platform now — which is a way around the missing omni checkpoint, not a replacement for its audio tower.

Screenshot of the OrcaRouter models catalogue showing a filter rail for input modalities, context length, input price, status, access tier, series, supported parameters, privacy and inference region, with a header reading 229 models across 17 providers on one API key and one bill.

Which narrowness you can live with

If you need footage this quarter and one reference image is enough, HiDream-O1-Video-1.0 is the only one of the two with a callable endpoint, and its constraints are documented well enough to integrate against in an afternoon. Budget $5.80 a minute, plan for ten seconds or the model's choice, and read the subtask status rather than the terminal flag.

If you need to know what is in footage you already own — including what is said over it — WeMM-Video-4B is the only one of the two that can answer, and the price of admission is a patched inference engine, a 10.45 GiB checkpoint, and a 16-frame sampling budget you should tune before you trust a long-video result.

What neither gives you is the thing a comparison page usually promises: a shared task on which to rank them. They are not two answers to one question. They are two models whose public surfaces are both smaller than their papers, and the useful engineering judgement is which of those gaps you can close with your own hardware.

Compared in this article1

Detected from this article