Hero title card reading 'WeMM-Video-9B vs Wan 3.0', subtitled 'One sells the seconds. One reads them back.', with three stat chips reading 10.49B open weights, 8s view by default and 30s clip ceiling, and a footer saying WeMM-Video-9B figures come from Tencent's repository and arXiv 2611.26409 while Wan 3.0 figures come from Alibaba and Artificial Analysis.
Guides & Insights

WeMM-Video-9B vs Wan 3.0: Seconds Sold, Seconds Watched

Author

Alistair Wren

Date Published

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

Wan 3.0 sells seconds of video. WeMM-Video-9B reads them back. That is the whole comparison, and it is the reason a dimension-by-dimension spec fight between these two would be close to meaningless: one is a generation model that returns an MP4, the other is an omni-modal understanding model that returns a sentence. They are not rivals. They are the two halves of one pipeline — produce a clip, then check it — and the interesting question is not which wins but where each one stops being usable.

Only one of them is news. Tencent uploaded WeMM-Video-9B and eighteen sibling repositories in a five-minute window starting at 08:53:50 UTC on 17 November 2026, published the technical report as arXiv 2611.26409 on 19 November, and has not announced any of it: no blog post, no launch page, no paid tier. At the time of writing the 9B repository shows zero downloads, the collection shows a handful of likes, and no inference provider offers the model. Wan 3.0 is the opposite case entirely — Alibaba put it into public beta on 6 August 2026 and launched it officially on 24 August, eighty-eight days before this piece, and it currently sits at the top of two independent leaderboards. Everything below about Wan 3.0 is a launched, priced, independently scored product. Everything about WeMM-Video-9B comes from either its own repository or the authors' PDF, and those are the only two places any number about it exists.

What each one actually returns

The specification lines are worth putting side by side precisely because they describe different jobs.

• Output — WeMM-Video-9B returns text, in Chinese or English. Wan 3.0 returns a generated clip of up to 30 seconds with native audio.

• Input — the 9B takes video frames with their original soundtrack, silent video, standalone audio, images and text, and can reason over all of them at once. Wan 3.0 takes text, image, video and audio prompts, plus, uniquely in its family, documents: doc, xls, ppt, pdf, md, txt, up to 100 MB or 50 pages per file, converted straight into a video.

• Weights — WeMM-Video-9B ships Apache 2.0 weights on Hugging Face in BF16 plus four quantised builds (FP8 Block128, AWQ W4A16, GGUF Q4_K_M, NVFP4 W4A16). Wan 3.0 has no published weights; it is a metered, closed API.

• Price — the 9B has no hosted price, because nothing hosts it; your cost is a GPU, a patched runtime and your own time. Wan 3.0 bills per second of output: 0.3, 0.6 and 1.2 yuan per second at 480p, 720p and 1080p — roughly $0.05, $0.10 and $0.20 — which Artificial Analysis normalises to $12.00 per minute on its leaderboards.

• Independent scores — none for WeMM-Video-9B. Artificial Analysis lists no WeMM model on any board. Wan 3.0 is first on the Artificial Analysis text-to-video board at Elo 1,176 (95% interval 1,163 to 1,189, 8,243 samples) and first on its video-editing board at 1,197 (1,187 to 1,207, 9,919 samples), and twelfth on image-to-video at 1,109.

Two-column scoreboard titled 'WeMM-Video-9B vs Wan 3.0 — the scoreboard'. Left column WeMM-Video-9B: output text in Chinese or English, weights Apache 2.0 open now, price no hosted API, independent Elo none, context declared 262,144, shipped video view 8s at 2 fps. Right column Wan 3.0: output video with audio, weights closed API only, price 0.3-1.2 yuan per second, independent Elo 1,176 text-to-video, context declared not published, shipped clip length 30s single run. Footer states WeMM figures are vendor-reported and unaudited, and the Wan 3.0 Elo is per Artificial Analysis, November 2026.

The price gap is not a price gap

Twelve dollars a minute is the number people will quote, and it is the wrong number to compare against anything about the 9B, because the two bills measure different quantities. Wan 3.0 charges for duration produced. WeMM-Video-9B has no per-unit charge at all — it charges in hardware you already own, and the honest way to price it is to work out how much of a card it occupies and how many clips you can push through per hour.

The occupancy figure is measurable from the repository. The BF16 checkpoint is 19.54 GiB of weight files and 10,488,603,384 parameters across 1,305 tensors — 10.49 billion, not nine, because the "9B" in the name refers to the Qwen3.5-9B language backbone rather than the whole stack. The AWQ W4A16 build comes in at 12.86 GiB. That is a 34% cut rather than the 75% a four-bit recipe implies, and the reason is visible in the exporter manifests: only the dense feed-forward matrices are compressed, while the vision tower, the Whisper-based audio encoder, the embedding, the output head and every attention weight keep their original precision. About 5.7 billion parameters stay at two bytes each, which sets a floor no four-bit recipe can go under. Anyone sizing a deployment from the name will be roughly 1.5 GiB and one precision tier off.

So the cost comparison that actually matters is not dollars against dollars. It is: one API call at $0.20 per second of finished footage, against a 13-to-20 GiB resident model, a patched serving stack and whatever your GPU-hour costs. For a team already running inference, the second can be cheaper. For a team that wants a clip by Friday, the first is the only one that exists.

Where the 9B lands against a 30B

The one place WeMM-Video-9B has hard numbers is the authors' evaluation, and the headline claim in the paper is about its bigger sibling — that the 35B-A3B model sets a new open-source state of the art on omni-modal video and closes on Gemini-3.0-Flash. The 9B gets a quieter claim: it outperforms Qwen3-Omni-30B-A3B-Instruct, a model roughly three times its size. In Thinking mode, on the joint audio-visual table, it does.

• Video-MME with audio — 85.1 for the 9B, 77.3 for Qwen3-Omni-30B-A3B-Instruct, 69.2 for MiniCPM-o-4.5-9B.

• Video-MME-v2 with audio — 51.6 against 36.3, a 15.3-point gap.

• Reasoning-flavoured rows — SciVideoBench 64.2 against 49.9, MMOU 71.7 against 60.5, EvaOmni 50.7 against 26.1, OmniReasoningBench 43.9 against 24.1.

• Where it loses — FutureOmni 53.7 against 55.1, and JointAVBench is a tie at 64.6.

The more interesting comparison is internal, and it is the one nobody quoting the paper mentions. Run the 9B against the 35B-A3B flagship on the same 17 rows and the gap is small: 85.1 against 86.4 on Video-MME with audio, 51.6 against 54.9 on Video-MME-v2, 71.7 against 74.5 on MMOU. On one row the smaller model wins outright — SVHalluc 86.8 against 80.3. A model carrying a fraction of the flagship's active parameters is within 1.3 points on the headline benchmark. That is the strongest argument in the repository for the 9B specifically, and it is an argument about deployment cost, not about capability.

The 9B is also not a uniform upgrade over the backbone it inherits. Comparing it with plain Qwen3.5-9B on the silent-video table shows a model tuned toward when and where something happened, at a measurable price elsewhere: F-16-NBA 88.7 against 75.8, TimeLens-QVHighlights 66.8 against 64.3, TOMATO 48.5 against 45.0 — but MLVU 82.4 against 84.4, LVBench 68.2 against 70.0, VSI-Bench 54.5 against 58.3, MVBench 71.4 against 74.4. If your task is long-video summarisation, the audio-visual machinery has cost you something the backbone already did better.

Eight seconds of coverage, by default

This is where the two models become directly comparable, because it is a limit on duration — and the numbers are not close.

Wan 3.0 generates up to 30 seconds in a single run, twice the 15-second ceiling of Wan 2.7, with a video-extension tool for chaining longer narratives. WeMM-Video-9B, as shipped, looks at eight.

The gap is not in the checkpoint. The 9B declares 262,144 position embeddings in its text config, which is a very large budget. It is in the recipes. The processor defaults to 2 fps with a minimum of four frames, a maximum of 128 frames and 100,663,296 total pixels, and the vLLM command Tencent publishes overrides all four to 2 fps, 16 frames and 262,144 pixels. At 2 fps, sixteen frames is eight seconds of coverage taken from the middle of a clip. Every serving command in the repository also passes --max-model-len 8192 or --context-length 8192, and nothing in the repository explains how to reach the larger declared figure.

• Frame budget — shipped default 16 frames at 2 fps, i.e. about 8 seconds, against Wan 3.0's 30-second single generation.

• Audio cost — the audio stack emits 10 tokens per second, so an 8,192-token budget is roughly thirteen minutes of audio and nothing else; a clip with a soundtrack spends that same budget on frames and audio together.

• Per-request media — four images, two videos and two audio items, per the shipped limits in both the vLLM and SGLang recipes.

Nobody outside WeChat Vision has published which sampling the benchmark runs used. If you are reproducing a number from the table above, that is the first unknown to pin down, and the shipped 16-frame default is the wrong place to start. But the ceiling is real: for anything longer than a take, the 9B is a spot-checker, not a reviewer, unless you raise the frame cap and pay for it in context.

Screenshot of the Artificial Analysis text-to-video leaderboard captured in headless Chromium at 1440 pixels wide, showing Wan 3.0 ranked first with an Elo of 1,176 above the rest of the board, with no WeMM model listed anywhere in the ranking.

The number nobody has measured yet

Three facts sit between this model and any confident recommendation, and they are worth stating flatly rather than burying.

The first is that every score above is vendor-reported. There is no independent evaluation of WeMM-Video-9B anywhere — no arena Elo, no third-party reproduction, no leaderboard entry. The tables come from the same authors who trained the model, on 120-plus benchmarks of their own selection, and several of those benchmarks use different judges: the paper's evaluation draws on GPT-6.1-Sol, GPT-4o and Qwen3-235B-A22B depending on the task. Cross-model comparisons inside the paper are therefore comparisons of numbers produced under different graders, which is a caveat the paper states and which anyone quoting the table tends to drop.

The second is a mode split that makes two rows non-comparable. The omni-modal video table runs the WeMM models in Thinking mode; the audio table runs them in Instruct mode, on the authors' judgement that standard audio benchmarks are direct perceptual tasks. That is defensible, and it also means a WeMM row in one table cannot be lined up against a rival row in the other.

The third is distribution. Zero downloads is not evidence of quality either way, but it does mean there is no body of practitioner reports — no failure modes, no "it hallucinates on music videos", no thread of people discovering the frame cap the hard way. Wan 3.0 has had eighty-eight days of that, plus launch integrations in ComfyUI, Leonardo.Ai and Envato. On the question "what actually goes wrong in production", one of these models has an answer and the other has a repository.

Screenshot of the Hugging Face repository page for tencent/WeMM-Video-9B, showing the model card at the top of the page and the repository tab bar, for a repository with no linked launch announcement.

Running the pair today

If you want both, they are installed in completely different ways, and neither is on a routing platform.

Wan 3.0 is a metered API call. It is reachable through Alibaba's own consoles and through third-party creative tools, it bills per second at 0.3, 0.6 and 1.2 yuan for 480p, 720p and 1080p, and the only real procurement question is which resolution your review step needs, because that is the multiplier.

WeMM-Video-9B is a self-host job. You need Python 3.12, an NVIDIA GPU on a CUDA 12.9 driver, FFmpeg and the TorchCodec shared libraries, then the repository's runtime patch: patched vLLM 0.18.0, or 0.31.0 for BF16, or patched SGLang 0.5.16 or 0.5.19, pinned to the exact PyTorch, Transformers and CUDA combination the patch manifest names for your engine version. Loading additionally requires trust_remote_code, because the model ships custom modelling and processor code. Tencent validated that patched vLLM loads the checkpoint and completes text, image, audio and video-with-soundtrack generation checks on an H20, which is a load-and-run validation, not an accuracy claim. The GGUF route needs the native runtime built from source.

OrcaRouter does not serve WeMM-Video-9B, and its catalogue lists no WeMM checkpoint at all, so there is no hosted price to quote and no endpoint to call. What the platform does carry is the vision half of the same job — Qwen3-VL-8B-Thinking and Qwen3-VL-235B-A22B-Thinking both take images and video frames — behind one OpenAI-compatible key alongside 200+ other models, billed at provider list price passed through at 0% markup. Neither of them hears anything, so neither substitutes for the audio-visual half. If the plan is to try a self-hosted 9B without putting a production path on it, routing it behind a failover rule is the cheap way to find out whether the patched runtime holds at your concurrency before you commit a card to it.

The pairing itself is a workflow, not a benchmark. Generate with Wan 3.0, then hand the clip to a model that returns a description instead of pixels: does the scene match the prompt, is the speech intelligible, does the audio land on the action. Nothing here measures whether WeMM-Video-9B is good at that job against Wan 3.0's own outputs specifically — nobody has run it. The eight-second frame budget is the constraint to test first, because if your clips are thirty seconds, the reviewer is looking at a quarter of each one.

What would change the answer

Three datable things, in order of how much they would move the recommendation.

An independent score for WeMM-Video-9B. A single Artificial Analysis entry, or one arena result from a party that does not share an employer with the authors, would convert the table above from a claim into a measurement. Until then the honest reading is that a 10.49B checkpoint with a 9B backbone scores close to its own 35B-A3B flagship on Tencent's own benchmarks, and that is all that can be said.

A hosted endpoint. The model needs a patched runtime, custom code and a 13-to-20 GiB resident footprint; the day a provider offers it behind an API, the price comparison in this article stops being a category error and becomes arithmetic.

A recipe that uses the declared context. The checkpoint already claims 262,144 positions. If Tencent ships a serving profile that raises the frame cap and actually exercises it, the eight-second default becomes a config detail rather than a ceiling — and the 9B stops being a spot-checker for short takes.

For now the split is clean. Pick Wan 3.0 when you need footage; it is launched, priced, top of two independent boards and the only one of the two you can call this afternoon. Pick WeMM-Video-9B when you need a model that hears the soundtrack as well as it sees the frames, you have a card to spare, and you are willing to be the first person to find out whether the eight-second recipe is a default or a limit. The two are not competing for the same decision.

Compared in this article2

Detected from this article · Benchmarks: Artificial Analysis · updated daily