Title card reading 'WeMM-Video-4B vs FastFrames Video Generation' with the subtitle 'A silent seam', two panels split by a vs disc, the left listing 5.61B, video and audio in, text out, 1.07B of it is ears, 16 frames maximum and Apache 2.0, the right listing 1.3B and 14B, text in, video out, 81 frames at 832 by 480, 36 ms per frame on one B200 and Apple AMLR, and a footer reading 'WeMM figures vendor-reported; FastFrames figures per the authors.'
Engineering & Research

WeMM-Video-4B vs FastFrames Video Generation: A Silent Seam

Author

jinhao.song

Date Published

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

FastFrames Video Generation arrived on 16 November 2026 and WeMM-Video-4B on 17 November, both as weight downloads with a paper attached and neither with a launch post worth the name. That coincidence is the only reason to read them against each other, because they do not do the same job: FastFrames turns a sentence into 81 frames of silent video, and WeMM-Video-4B turns video into a description, including a description of what is being said over it. The interesting question is not which is better. It is what happens when you put one behind the other — and the answer is that the seam between them is silent, which quietly disables the most expensive part of the WeMM checkpoint.

Everything below is either measurable in the two repositories or reported by the authors of the two papers. Neither model has been run by anyone who did not write its evaluation section, and neither has a hosted price, an arena row or a third-party reproduction. Where a figure comes from a vendor table it is labelled as such.

What FastFrames puts on disk

FastFrames is an autoregressive video diffusion model distilled from Wan2.1, published by Apple and Cornell as arXiv 2611.23226 on 17 November, with weights going up a day earlier. It comes in two sizes — FastFrames-S at 1.3 billion parameters and FastFrames-L at 14 billion — and the thesis is latency rather than fidelity. On a single B200 the small model produces a frame in 36 milliseconds, which is faster than the clip plays back: a 5.06-second generation takes roughly 2.9 seconds. The large model manages 45 milliseconds per frame spread across eight B200s, and degrades to 158 milliseconds on one GPU, which is slower than real time. Apple also reports a 25 to 50 percent penalty on H100s instead of B200s.

The output format is fixed and narrow. 832×480, 16 frames per second, 81 frames per generation — a hair over five seconds — driven by a text prompt alone. There is no image conditioning, no keyframe input, no continuation mode and no camera control in the released configuration. There is also no audio: no audio decoder, no audio latent, no soundtrack. The authors report VBench totals of 84.23 for the small model and 84.44 for the large one at 480p on matched prompts and seeds, which is a respectable score for a model whose selling point is that it is fast, and it is the only quality number in the release.

Where 1.07 billion parameters go

WeMM-Video-4B is the middle child of a four-size family from T​encent's WeChat Vision group — 2B, 4B, 9B dense and a 35B-A3B mixture of experts — all uploaded inside a nineteen-minute window on 17 November, with the technical report, arXiv 2611.26409, following on 19 November. The name understates it: the safetensors metadata records 5,611,746,312 parameters, or 10.45 GiB in BF16, because "4B" refers to the Qwen3.5-4B language backbone rather than the checkpoint.

Reading the tensor index explains why. Of 1,268 tensors, the language model carries 4.21 billion parameters, the vision tower 0.33 billion, and the audio stack 1.07 billion. Nearly a fifth of this checkpoint is a Whisper-style encoder-decoder — 32 encoder and 32 decoder layers at width 1,280, with a window-level Q-Former that compresses 0.1-second windows into 10 audio tokens per second, fed at 16 kHz in 30-second chunks. That stack is initialised from Qwen2.5-Omni.

Screenshot of the Hugging Face model page for tencent/WeMM-Video-4B showing an Image-Text-to-Text pipeline tag, Safetensors and custom-code tags, an Apache 2.0 licence, a BF16 block reading 10.45 GiB of weight files, a size panel reading 6B params and 10.5 GB, and a Quantizations section listing four sibling repositories.

• FastFrames Video Generation — text in, frames out, no sound anywhere in the pipeline.

• WeMM-Video-4B — frames and sound in, text out, with the sound handled by a dedicated 1.07B tower rather than inferred from pixels.

• Overlap — none at the model level. Both are Apache-or-Apple-licensed weight bundles with no endpoint behind them, and both expect you to own the accelerator.

The only clip FastFrames makes is silent

Follow that through and the composition problem becomes concrete. If you wire WeMM-Video-4B behind FastFrames Video Generation — render a shot, then describe it so you can index, filter or QA it — every clip you hand the describer has no audio track. The 1.07-billion-parameter audio tower has nothing to listen to. The Q-Former emits nothing, and the fusion stage reduces to the vision path.

That is not a defect in either model; it is a fact about the pipeline, and it has a measurable consequence, because the WeMM paper reports silent-video benchmarks separately. On vision alone, WeMM-Video-4B scores 75.7 on Video-MME, 65.3 on LVBench, 74.8 on LongVideoBench and 74.2 on EgoSchema. Its joint audio-visual numbers — 83.9 on Video-MME with audio, 82.5 on Daily-Omni — are the ones that justify the file size, and they are exactly the ones a FastFrames-fed pipeline cannot reach.

So the honest framing of this pairing is that FastFrames exercises about 80 percent of the WeMM checkpoint. If your goal is a captioning loop over generated footage, a vision-only model half the size would do the same work; if your goal is understanding footage that has sound in it, FastFrames cannot supply the input. The two tools meet at a point where one of them is half-idle.

Screenshot of the Tencent/WeMM-Video GitHub repository showing six stars and zero forks, a README describing an omni-modal video understanding family from WeChat Vision, and a file tree containing patch, native, examples, licenses and a THIRD_PARTY_NOTICES file.

Two speed numbers, no shared units

Both releases lead with a latency claim and the claims are not commensurable, which is the most common way this comparison gets mangled on the internet.

FastFrames' 36 milliseconds is per output frame on one B200, measured against playback. WeMM-Video-4B does not produce frames; its cost is the cost of decoding text over a capped media budget. That budget is tighter than the processor defaults suggest. The vLLM recipe in the repository sets max_frames 16 against a processor default of 128, fps 2, and a total pixel cap of 262,144, with a context length of 8,192 tokens. Audio spends 10 tokens for every second before a single frame is counted.

The practical reading: a 20-second clip and a 5-second clip both arrive at the describer as sixteen frames, and a clip with a soundtrack spends its context on two streams at once. FastFrames gives you motion at 16 frames per second of output; WeMM-Video-4B sees a slideshow and infers the motion. Neither number tells you anything about the other, and a page that puts "36 ms" next to "8K context" is comparing a render budget to a reading budget.

Two-column scoreboard headed 'WeMM-Video-4B vs FastFrames Video Generation', the left column listing direction, 5.61B total parameters, a 1.07B audio stack, an 8,192-token and 16-frame budget, Apache 2.0 and no independent benchmark, the right listing text in and video out, 1.3B small and 14B large, no audio, 81 frames at 832 by 480, Apple AMLR and no independent benchmark, with a footer reading 'WeMM figures vendor-reported; FastFrames figures per the authors and not independently reproduced.'

Licences decide more than scores here

For a comparison this thin on independent evidence, the licence line carries more decision weight than any benchmark in either paper, and the two sit at opposite ends of the spectrum.

• FastFrames Video Generation — Apple's AMLR terms on the weights, with the LingBot-World v2 demo components under CC BY-NC-SA 4.0. The "NC" is the constraint that matters: open weights, not open for commercial use.

• WeMM-Video-4B — Apache 2.0 across the collection, with a T​encent licence file stating that third-party components "remain subject to their respective original terms". The carve-out is not itemised, which is a real gap rather than a formality given that the audio stack descends from Qwen2.5-Omni and the backbone from Q​wen3.5, both Apache 2.0.

• Both — full training-stage checkpoints and an evaluation pipeline in the repository, which is unusually reproducible for releases of this kind and is the best available substitute for the independent evaluation neither has.

Wiring a describer behind a renderer

If you build the loop anyway, the engineering question is where the text model goes, because there are three text hops and only one of them is optional.

• Prompt expansion — a shot note becomes a FastFrames prompt. Necessary, because a 1.3B distilled diffusion model has no room to interpret a vague instruction.

• Description — the rendered clip goes to WeMM-Video-4B and comes back as prose, which is the step that makes the footage searchable.

• Selection — the prose decides whether the clip is kept, which is the step that has to be cheap, because you pay for it on every candidate rather than every survivor.

Those two outer hops are ordinary text completions, and they are the part of this stack OrcaRouter actually covers: one OpenAI-compatible key across 200-plus models, provider list prices passed through with 0% markup so a vendor price change lands the same day, and automatic failover so a rate limit in the middle of a thousand-clip batch becomes a retry instead of a hole. Neither FastFrames Video Generation nor WeMM-Video-4B is on the platform — both are weight downloads, and no routing layer turns a checkpoint into an endpoint. What the router removes is the integration work around the two models you do self-host.

What is still unconfirmed

• Independent quality — no Artificial Analysis row, no arena entry, no third-party reproduction for either model. VBench 84.44 and Video-MME 83.9 are author-reported and were not run by anyone else.

• Evaluation conditions — the WeMM paper does not state the frame count, pixel budget or context length used for its video benchmarks, and the shipped recipe uses 16 frames where the processor defaults to 128. That gap is large enough to matter on the long-video rows.

• Price — neither has one. No hosted endpoint, no per-second rate, no partner list. The cost is the GPU you own, plus the patched vLLM or SGLang overlay the WeMM repository requires, since neither engine supports the architecture upstream.

• Commercial usability on the Apple side — AMLR plus a non-commercial demo licence means the renderer, not the describer, is the part of this pipeline a legal review will stop.

The question that picks one

Ask what your artefact is. If it is a file, only FastFrames can produce it, and you plan around 480p, five seconds, silence and a non-commercial licence. If it is a claim about footage you already have — what happens, who speaks, what is audible — only WeMM-Video-4B can produce it, and you plan around sixteen sampled frames and eight thousand tokens.

If it is both, you are building a two-model system whose seam is silent, and the cheapest correct version of that system does not use a 5.61B omni checkpoint to caption silent output. Use a vision model for the rendered clips and keep WeMM-Video-4B for source footage that carries sound. That is a less elegant architecture than one model for everything, and it is the one the numbers support.

What neither release provides yet is the thing that would settle it: a run by somebody who did not write the paper. Two labs shipped serious video work in the same week without an announcement, which is the more interesting fact about November 2026 than any single benchmark cell.