Title card reading 'WeMM-Video-4B vs FLUX 3 Video' with the subtitle 'Read the clip or make one', two panels split by a vs disc, the left badged 'open weights' and listing $0.00 per minute, 5.61B parameters and an 8,192-token context, the right badged 'metered' and listing $17.40 per minute, up to 20 seconds and 10th on AA-Video-T2V v2.1, with a footer reading 'Artificial Analysis figures read 20 November 2026.'
Guides & Insights

WeMM-Video-4B vs FLUX 3 Video: Read the Clip or Make One

Author

jinhao.song

Date Published

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

The fastest way to waste an afternoon is to compare WeMM-Video-4B with FLUX 3 Video on quality. One of them has a number and the other does not, and the number is not a quality score — it is a price. FLUX 3 Video, from Black Forest Labs, went to general availability on 4 August 2026 with a published per-minute rate and rows on both Artificial Analysis video boards. WeMM-Video-4B, from T​encent's WeChat Vision group, appeared on 17 November 2026 as ten and a half gigabytes of Apache-2.0 weights with no endpoint, no price and no leaderboard entry at all. What you are choosing between is a metered generator and a free describer, and the honest comparison is about which of those two things your pipeline is missing.

The rest of this piece keeps three kinds of fact apart: what the repositories and the two boards state, what the vendors claim in their own papers and documentation, and what nobody outside either lab has verified. As of 20 November 2026, that third category is most of the interesting material.

One is on two boards, the other on none

Artificial Analysis scores FLUX 3 Video twice, and the two scores live on separate scales, so neither transfers to the other.

• AA-Video-T2V v2.1 — FLUX 3 Video 10th, Elo 1,137 ±13 over 7,406 votes, listed at $17.40 per minute. The larger FLUX 3 Video Max sits 7th at 1,151 ±14 over 4,708 votes at $8.70 per minute, which is the unusual case of the bigger model being the cheaper one on the board.

• AA-Video-I2V v2.1 — FLUX 3 Video 16th at 1,069 ±14 over 6,838 votes, same $17.40. FLUX 3 Video Max is 17th at 1,068 ±14 over 6,848 votes at $8.70.

• WeMM-Video-4B — no row on either board, and no row on the video-editing board. It is not a missing measurement; it is not enterable. The AA video boards rank generated footage, and WeMM-Video-4B generates none.

Read the intervals rather than the ranks. Sixty-eight points separate FLUX 3 Video's two placements across the text-to-video and image-to-video boards — five intervals' worth — which is the useful fact about a model scored on two instruments that do not convert into each other: it is a competent generator whose position depends on which task you put to it. That is a real, if modest, independent measurement — and it is more than the other side of this matchup has.

What $17.40 a minute buys

FLUX 3 Video is a multimodal generator with a broader input list than any single board row shows. Text prompts, a still image as first frame, video for video-to-video and continuation, and keyframes for controlled transitions. Duration runs to 20 seconds in one generation, with 5 to 20 seconds for text-to-video and image-to-video and 5 to 15 for continuation. Resolution is sold in pixel bands rather than aspect ratios — HD at about one megapixel per frame, FHD at two, QHD/2K at four, UHD/4K at eight. Native audio with multilingual dialogue, synchronized speech, sound effects and ambience is optional and included in the per-second price.

• Draft HD — $0.06 per second. A five-second test clip is thirty cents.

• HD / 720p — $0.17 per second, or about $0.85 for five seconds.

• FHD / 1080p — $0.29 per second. QHD/2K is $0.40 and UHD/4K is $0.80.

• Video-to-video and continuation — $0.12 draft, $0.41 to $0.43 HD, $0.53 to $0.54 FHD, $0.65 at 2K and $0.95 at 4K.

The board's $17.40 per minute is a normalised 1080p figure, which is why it does not divide neatly into that card. It is a comparison unit, not a receipt. Several third-party summaries also report the 2K and 4K bands as rolling out rather than universally available, so treat the top of the card as an account-level question.

What WeMM-Video-4B gives you for nothing

The other side of the matchup costs nothing per call and cannot be bought at all. WeMM-Video-4B is 5,611,746,312 parameters — 10.45 GiB in BF16 — in a repository named for its Qwen3.5-4B language backbone rather than for its actual size. The surplus is the point: 4.21 billion parameters of language model, 0.33 billion of vision tower, and 1.07 billion of Whisper-style audio stack with a Q-Former emitting 10 audio tokens per second from 16 kHz audio in 30-second chunks.

It takes text, images, video and video with its original soundtrack, and returns text. On the authors' joint audio-visual table it beats Qwen3-Omni-30B-A3B-Instruct on Video-MME with audio, 83.9 to 77.3, and on Video-MME v2, 48.1 to 36.3. On silent video it leads LVBench 65.3 to 50.2 and LongVideoBench 74.8 to 63.5. Against Qwen3.5-Omni-Flash it loses the audio benchmarks — MMAU 75.2 to 80.4, MMAR 67.3 to 74.0 — and trails on OCRBench 85.7 to 89.1 and GPQA 73.2 to 76.4.

Every one of those is T​encent's own measurement. The paper says it reports official results where available and its own evaluations otherwise, and no third party has reproduced any row. Frame counts, pixel budgets and context length for the benchmark runs are not stated — which matters because the shipped vLLM recipe caps video at 16 frames and 262,144 total pixels against a processor default of 128 frames.

Two-column contrast headed 'WeMM-Video-4B vs FLUX 3 Video', the left listing describe footage, 5.61B parameters, text image video and audio inputs, text output, no price and self-hosting, and no independent score, the right listing generate footage, parameters not published, text image video and keyframe inputs, up to 20 seconds with native audio, $0.06 to $0.80 per second and 10th at Elo 1,137, with a footer reading 'Artificial Analysis figures read 20 November 2026; WeMM figures vendor-reported.'

Two clocks that never meet

Neither model can stand in for the other, and the constraints that make that true are both budgets rather than capabilities.

• Output budget — FLUX 3 Video produces up to 20 seconds of footage per call at up to 4K with audio. WeMM-Video-4B produces no frames at all, so its output budget is a token budget: 8,192 tokens per request as shipped, with audio consuming 10 tokens per second before frames are counted.

• Input budget — WeMM-Video-4B accepts four images, two videos and two audio items per request, at 2 frames per second capped to 16 frames. FLUX 3 Video accepts a prompt, one first frame, one reference video or a keyframe set, and no audio input of any kind.

• Cost shape — FLUX 3 Video charges per second of output, so cost scales with ambition. WeMM-Video-4B charges per GPU-hour of your own hardware, so cost scales with batch size and does not care how long the source clip is, only how many frames survive sampling.

• Licence — Apache 2.0 on the WeMM collection, with an un-itemised carve-out for third-party components; a hosted commercial API on the FLUX side with no weights published.

Put simply: one is a per-second expense with a scoreboard, the other is a fixed capital cost with a paper. A team that needs a finished shot cannot use WeMM-Video-4B, and a team that needs a searchable index of ten thousand existing clips cannot use FLUX 3 Video — not because it is expensive, but because it has no way to answer a question about footage it did not make.

Audio is where the comparison bends

FLUX 3 Video generates audio. WeMM-Video-4B understands audio. Those are different capabilities that sound similar in a feature list, and the difference decides the pipeline.

FLUX's audio is a generation output: multilingual dialogue, synchronized speech, sound effects and ambience, optional and bundled into the per-second price. What Black Forest Labs does not publish is a track count, so if you need dialogue, music and effects as separable stems, neither the rate card nor the documentation tells you whether you will get them. Budget a second pass in a DAW.

WeMM's audio is an input path: a 1.07-billion-parameter tower that turns a soundtrack into 10 tokens per second and fuses them with visual tokens in chronological order, with explicit textual timestamps. That is what the 83.9 on Video-MME with audio measures and what the 75.2 on MMAU measures — how well it hears, not how well it speaks.

So a workflow that writes a scene, renders it with sound, and then needs to know whether the dialogue actually landed requires both halves, and neither model covers the other's half. That is the one place in this matchup where the two are genuinely complementary rather than merely different.

The routing question, stated honestly

OrcaRouter serves neither model. There is no WeMM-Video-4B endpoint — the platform's catalogue has no WeMM checkpoint at all — and no FLUX 3 Video route either. FLUX 3 Video is available through Black Forest Labs' own API and several third-party platforms; WeMM-Video-4B is available as a download, with a patched vLLM or SGLang overlay, because neither engine supports the architecture upstream.

What the platform does cover is the text layer that surrounds either workflow, and on this matchup the useful piece is model fusion rather than plain routing. The decision "is this clip worth keeping" is a judgement call, and a fused call that puts a cheap classifier and an expensive reasoner on the same prompt can make it at a fraction of the cost of running the expensive one on every candidate. That runs on one OpenAI-compatible key across 200-plus models, at provider list prices with 0% markup, so when a vendor cuts a price the cut is live the same day rather than at the next billing cycle. For the vision-only half of the WeMM work specifically, Qwen3-VL-235B-A22B-Thinking and Qwen3-VL-8B-Thinking both take images and video frames and are on the platform today — which is a route around the missing omni model, not a substitute for its audio tower.

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.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 Max at rank 7 on Elo 1,151 and FLUX 3 Video at rank 10 on Elo 1,137.

What would actually settle it

Three developments would move this comparison off the vendor-reported footing it currently sits on.

First, a WeMM-Video-4B entry on any independent board. There is none, and the reason is structural rather than conspiratorial: the AA video boards rank generated footage, and an understanding model cannot be entered. An omni-modal video-understanding arena is the missing instrument, and until one exists every WeMM number is T​encent's word.

Second, a FLUX 3 Video release of weights. Black Forest Labs has an open-release reputation and this checkpoint has no licence for download, so the comparison is currently between an open model and a closed one, which is a difference no benchmark row captures.

Third, a hosted price for the WeMM family. If a compatible host starts serving these weights at provider list price with 0% markup passed through, the cost side of this article changes shape within a week. That is a condition, not a current fact, and nothing here should be read as a claim that we host it.

Which one you actually need

If the deliverable is footage, FLUX 3 Video is the only one of the two that can produce it, and the decision reduces to whether $0.17 per second for 720p is acceptable against the alternative generators on the same board — which, at 10th on text-to-video, it is priced mid-field for.

If the deliverable is a statement about footage you already hold, WeMM-Video-4B is the only one of the two that can produce it, and the decision reduces to whether you can carry a source overlay for vLLM or SGLang and a 10.45 GiB checkpoint to get a 4B-class describer with unusually good long-video numbers.

The trap is treating the $17.40 and the $0.00 as the same quantity. One is a metered rate for a service; the other is the absence of a service. WeMM-Video-4B is free the way a research release is free — you pay in integration, in unvalidated GGUF bundles, and in the absence of anyone else's opinion about whether it works.

Compared in this article1

Detected from this article