Title card reading 'WeMM-Video-9B vs FLUX 3 Video' with the subtitle 'Who bills for what', two panels split by a vs disc: the left badged 'weights only' and listing 9B dense parameters, 19.54 GiB of BF16 shards and a 128-frame ceiling, the right badged 'metered' and listing Elo 1,137 on AA-Video-T2V v2.1 and $17.40 per minute.
Guides & Insights

WeMM-Video-9B vs FLUX 3 Video: Who Bills for What

Author

jinhao.song

Date Published

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

Two models with video in their name, and only one of them will ever hand you a video. FLUX 3 Video, from Black Forest Labs, went to general availability on 4 August 2026 and sells seconds of finished 1080p footage at a published rate. WeMM-Video-9B, from T​encent's WeChat Vision group, went up on 17 November 2026 as a 19.54 GiB Apache-2.0 checkpoint that reads footage and writes sentences about it. Nobody announced it — no blog post, no pricing page, no demo, no board entry — which is why the useful question here is not which model is better. It is which of the two invoices you are trying to reduce.

That reframing is not a dodge. A comparison between a generator and a describer is only honest if you say up front that they are not substitutes, and then go looking for the places where they actually touch: cost per unit of work, the ceiling on how much of a clip either one will look at, and how much of either claim has been checked by anyone who does not work at the company.

Two product categories, not two competitors

FLUX 3 Video takes a text prompt, or a still, or a source clip, and returns new footage. Black Forest Labs documents text-to-video, image-to-video, video-to-video, keyframe conditioning and continuation, output up to twenty seconds at 1080p SDR MP4, with synchronised audio generated in the same pass. It is not open weights. There is no checkpoint to download and no licence that lets you run it on your own hardware; you rent it.

WeMM-Video-9B takes frames plus the original soundtrack, plus standalone audio, plus images, plus text, and returns text in English or Chinese. It is Apache 2.0 with a third-party carve-out the model card does not itemise. There is a checkpoint, five BF16 shards of it, and no hosted endpoint that anyone has found.

• Direction — FLUX 3 Video: text or image in, video out. WeMM-Video-9B: video and audio in, text out.

• Parameters — FLUX 3 Video: not published. WeMM-Video-9B: 9B dense language backbone, 10,488,603,384 total across 1,305 tensors.

• Ceiling — FLUX 3 Video: twenty seconds of output. WeMM-Video-9B: sixteen sampled frames and 8,192 tokens of input, as shipped.

• Access — FLUX 3 Video: metered API. WeMM-Video-9B: download and self-host, with a vLLM or SGLang overlay.

Two-column contrast headed 'WeMM-Video-9B vs FLUX 3 Video', the left listing 10.49B parameters, 16 sampled frames in, text out and Apache 2.0, the right listing Elo 1,137 on AA-Video-T2V v2.1, up to 20 seconds and $17.40 per minute, with a footer reading 'Artificial Analysis figures read 20 November 2026; WeMM-Video-9B figures vendor-reported.'

Reading the $17.40

Artificial Analysis prices FLUX 3 Video at $17.40 per minute on both boards it appears on, and that consistency is worth pausing over, because it is the only number in this article that both sides of the market agree on.

The per-second card in the Black Forest Labs documentation is more granular and less flattering: $0.15 per second for HD, $0.24 for Full HD, $0.28 for QHD and $0.41 for UHD on text-or-image-to-video, with a separate and higher ladder for video-to-video at $0.21, $0.27, $0.31 and $0.45, and a Draft tier at $0.06. The model page FAQ still quotes $0.17, $0.29, $0.40 and $0.80 for the same four tiers. Two documents from the same vendor, both live, disagreeing by up to a factor of two at the top end. Budget from the per-second card and keep the FAQ as evidence that price pages rot.

This is where a routing layer earns its keep rather than decorating the piece. OrcaRouter passes provider list price through at 0% markup, so when a vendor revises a rate card — or publishes two that contradict each other — the number you are billed tracks the number they publish, not a reseller's cached copy. For a workload billed per second across four resolution tiers, that is the difference between a cost model that survives a price change and one that quietly drifts.

What the numbers on the board mean

FLUX 3 Video is ranked twice, on two instruments that do not convert into each other.

• AA-Video-T2V v2.1 — 10th at Elo 1,137 ±13 over 7,406 samples, July 2026, $17.40 per minute.

• AA-Video-I2V v2.1 — 16th at Elo 1,069 ±14 over 6,838 samples, same price. Its larger sibling FLUX 3 Video Max sits 7th on text-to-video at 1,151 and 17th on image-to-video at 1,068, at $8.70 per minute.

Sixty-eight points separate the text-to-video and image-to-video rows — five intervals' worth — which is the single most useful fact about a model scored on two boards: its rank is a statement about the task, not about the model. Neither board has a row for WeMM-Video-9B, and neither ever will, because both rank generated footage and WeMM-Video-9B generates none. That is not a low score. It is an empty cell.

Screenshot of the Artificial Analysis text-to-video leaderboard, AA-Video-T2V v2.1, showing ranked rows with Elo, confidence interval, sample counts, release month and per-minute pricing, with FLUX 3 Video at rank 10 on Elo 1,137 over 7,406 samples at $17.40 per minute and FLUX 3 Video Max at rank 7 on 1,151.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 cost per minute, including FLUX 3 Video at Elo 1,069 and FLUX 3 Video Max at 1,068.

What sixteen frames cannot see

The shipped configuration of WeMM-Video-9B is the part of the repository most likely to disappoint someone who skimmed the paper. The processor defaults allow 2 frames per second, a minimum of 4 frames, a maximum of 128 frames and a 100,663,296-pixel budget. The vLLM recipe that ships in the repo overrides all of it to 2 fps, 16 frames and 262,144 pixels, sampled around the midpoint of the clip. Sixteen frames at 2 fps is eight seconds of coverage, and the sampling starts in the middle, so the head and tail of a longer clip are simply not in the input.

Every serving command also passes an 8,192-token context length. At ten audio tokens per second the audio tower can hold roughly thirteen minutes of sound; the video side runs out far sooner. The architecture supports 262,144 position embeddings, so the ceiling is a recipe choice rather than a hard limit — but it is the choice the vendor shipped, and it is the one you get if you follow the instructions.

Against FLUX 3 Video's twenty-second output, that is a mismatch in both directions. Twenty seconds of generated footage is more than eight seconds of looking. If your pipeline generates and then describes, the describer is sampling a fraction of what the generator produced, and closing that gap is a configuration job you have to do yourself.

The evidence gap is asymmetric

FLUX 3 Video has two independent Elo rows and a price. What it does not have is a third-party check on the marketing: the FLUX 3 Max page claims "10-second video in 12 seconds," "22x faster than Seedance 2.5," and a tie for first place in image-to-video, with no named board, no Max price row on the per-second card, and no Max endpoint. Those are vendor claims and should be read as such.

WeMM-Video-9B has the opposite problem. Its benchmark table is detailed — Video-MME with audio 85.1 against Qwen3-Omni-30B-A3B-Instruct's 77.3, Daily-Omni 83.6 against 70.7, SciVideoBench 64.2 against 49.9, OmniReasoningBench 43.9 against 24.1, with a single loss on FutureOmni at 53.7 against 55.1 — and none of it has been reproduced outside the authors' own evaluation harness. The paper discloses this: "official results where available and our own evaluations otherwise." That sentence is doing a lot of work, and the zero download count on all nineteen WeMM-Video repositories three days after publication suggests nobody has started checking.

The decision, in one line each

Buy FLUX 3 Video if you need footage and $17.40 a minute is acceptable against a board where it ranks 10th on text-to-video and 16th on image-to-video. You get a documented API, a published price and an independent score, and you give up the ability to run it yourself.

Download WeMM-Video-9B if you need a description of footage you already have and can carry a 19.54 GiB checkpoint plus a patched runtime. You get a 9B describer with strong audio-visual numbers and an Apache-2.0 licence, and you give up independent verification, a serving contract, and any guarantee about how much of a long clip it will actually look at.

What you should not do is treat the $17.40 and the free download as the same line item. One is a rate. The other is a research release whose real cost shows up later, in integration hours and in the absence of a second opinion.