Title card reading 'WeMM-Video-4B vs Kandinsky 6.0 Video' with the subtitle 'Two hard caps', two panels split by a vs disc, the left subtitled 'reads clips' and listing 16 frames sampled in, 8,192 tokens and Apache 2.0 beside a small window icon captioned 'input cap', the right subtitled 'makes clips' and listing five seconds, 121 frames and 44.1 kHz audio beside a stopwatch icon captioned 'output cap', with a footer reading 'Both absent from every Artificial Analysis video board, 20 November 2026.'
Engineering & Research

WeMM-Video-4B vs Kandinsky 6.0 Video: Two Hard Caps

Author

jinhao.song

Date Published

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

Both Kandinsky 6.0 Video and WeMM-Video-4B are open-weight releases with permissive licences, both arrived in the last two months without a marketing campaign worth the name, and both are absent from every Artificial Analysis video board. That shared invisibility is where the resemblance ends, and what replaces it is a pair of ceilings that tell you more about each model than any benchmark row would. Kandinsky 6.0 Video makes five seconds of footage with synchronised audio and then stops — five seconds is the trained length, not a default. WeMM-Video-4B reads a clip through sixteen sampled frames and then stops, which is what its own serving recipe sets against a processor default of 128. One cap is on output duration, the other on input fidelity, and choosing between them is choosing which kind of blindness your pipeline can survive.

Neither model has been independently evaluated. Sber's numbers are Sber's, T​encent's are T​encent's, and the section headings below keep that boundary visible rather than burying it.

A five-second ceiling and a sixteen-frame window

Kandinsky 6.0 Video was released by Sber's Kandinsky Lab on 6 October 2026 under MIT, in two sizes: Pro at 29 billion parameters and Lite at 3 billion. It takes a text prompt or a reference image and returns a five-second clip with 44.1 kHz audio generated in the same pass, lip-sync included. The paper and the serving recipe are both built around 121 frames at 24 frames per second, and there is no extension mode in the release. Anything longer is a stitching problem you own.

WeMM-Video-4B is the middle child of a four-size family from T​encent's WeChat Vision group, uploaded on 17 November 2026 with the technical report following on 19 November. It is 5,611,746,312 parameters — 10.45 GiB in BF16 — because the "4B" label refers to its Qwen3.5-4B language backbone rather than the checkpoint. It accepts text, images, video and audio and returns text.

• Kandinsky 6.0 Video — output cap: five seconds, 121 frames, 24 fps, 44.1 kHz audio welded to the picture.

• WeMM-Video-4B — input cap: 2 frames per second sampled to a maximum of 16 frames, 262,144 total pixels, inside an 8,192-token context that audio spends at 10 tokens for every second.

• Both — the cap is in the shipped configuration, not the architecture, and neither repository explains what the model would do without it.

Read the two together and a workflow suggests itself that neither vendor documents: a five-second clip handed to a describer that samples sixteen frames is a clip sampled at roughly three frames per second, which is above the recipe's own two-per-second floor. That is the one place these two models fit together cleanly, and it is worth noticing precisely because it is accidental.

Licences are the easy part this time

For once the legal comparison is not the decision. Both releases are permissive and both are properly documented on this point, which is a relief after a year of Apple AMLR and non-commercial demo components.

• Kandinsky 6.0 Video — MIT on code, weights and the diffusers integration. The repository ships two ComfyUI nodes installable through ComfyUI Manager and two Hugging Face Spaces, one for the Pro distill checkpoint and one for super-resolution.

• WeMM-Video-4B — Apache 2.0 across the collection, with a T​encent licence file stating that third-party components remain under their 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 — commercial use is available, and neither vendor publishes a usage policy that narrows it further. If a legal review is your blocker, it will not be the licence line here; it will be the un-itemised carve-out on the T​encent side.

Deployment surface is where they part

The sharper difference is not licence but how much somebody else has already done for you, and the gap is wide enough to change a timeline.

Kandinsky 6.0 Video arrived with a diffusers bundle, a command line, two ComfyUI nodes and two Spaces on day one. Then an outside contributor merged pull request 8537 into vLLM-Omni, registering the pipeline as Kandinsky6TI2VAPipeline with video and audio latents denoised jointly by one transformer, weights loaded folder by folder from transformer/, vae/, text_encoder/, text_encoder_2/ and audio_vae/, and a serving recipe at recipes/Kandinsky/Kandinsky6-TI2VA.md. Reference-image conditioning is applied as a masked tail frame rather than a separate mode, so a single pipeline serves both text-to-video-and-audio and image-to-video-and-audio. Audio is on by default: a request with no audio parameters still returns synchronised speech, music or ambience.

WeMM-Video-4B has the opposite profile. Neither vLLM nor SGLang supports the architecture upstream, and a search of the vLLM issue tracker turns up no WeMM-Video request. What the repository ships instead is a patch directory — source overlays with SHA-256 manifests for vLLM 0.18.0 and 0.31.0 and 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, CUDA 12 NPP on the linker path for TorchCodec, and a C++ toolchain plus Ninja because SGLang compiles kernels on first use. The GGUF bundle targets a native runtime you compile from the repository rather than llama.cpp, and T​encent labels it exported_unvalidated.

Screenshot of vLLM-Omni pull request 8537, titled Kandinsky6 and merged on 6 October, describing a Kandinsky 6 TI2VA pipeline serving text-to-video-and-audio and image-to-video-and-audio, one DiT jointly denoising video and audio latents, weights loaded per Hub folder, audio on by default at 44.1 kHz and a reference image applied as a masked tail frame.

That asymmetry has a cost you can put in a schedule. Kandinsky 6.0 Video is a container that speaks HTTP today. WeMM-Video-4B is a day of toolchain work before your first token, and the only validation T​encent claims is that patched vLLM loaded it on an H20 and completed generation checks for text, image, audio and video with a soundtrack — no throughput, no latency, no memory footprint, no concurrency numbers.

Two-column contrast headed 'WeMM-Video-4B vs Kandinsky 6.0 Video', the left listing video and audio in with text out, 5.61B parameters, a 16-frame cap, patched vLLM or SGLang and Apache 2.0, the right listing text or image in with video and audio out, 29B Pro and 3B Lite, a five-second 121-frame cap, diffusers plus ComfyUI plus vLLM-Omni and MIT, with a footer reading 'Both are absent from every Artificial Analysis video board as of 20 November 2026.'Screenshot of the Hugging Face model page for kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers showing an Image-to-video pipeline tag, Diffusers, Safetensors, Kandinsky6TI2VAPipeline and video-generation tags, a model card line reading License: mit, and a gated-access notice asking the visitor to accept conditions before the files can be downloaded.

What the numbers say, and who said them

Artificial Analysis runs three video boards — text-to-video, image-to-video and video editing — each with its own Elo scale and its own vote pool. Read all three on 20 November 2026: Kandinsky 6.0 Video appears on none of them, and neither does WeMM-Video-4B. Neither absence is a judgement. Kandinsky 6.0 Video has no arena submission and the board's text-to-video ranking requires an API price for a normalised comparison, which an open checkpoint does not have; WeMM-Video-4B cannot be entered in a vote between two generated clips at all, because it generates none.

What is left is the vendors' own work, and it measures different things on different scales.

• Kandinsky 6.0 Video, quality — vendor-reported results on VABench and a lab-run human evaluation against Kling 2.6, Veo 3.1 Fast, MiniMax H3 and Seedance 2.0. Sber also reports, in the paper's ablation, that the distilled checkpoint it expects you to serve is preferred or tied on the majority of criteria at a mean preference of 51 percent to 49 percent — parity, at the speed you want.

• Kandinsky 6.0 Video, audio — two word-error-rate numbers in the paper that are not comparable. The 47 percent reduction accompanying the reinforcement-learning stage is scored with Whisper-large-v3, one of the RL reward models; the WER reported alongside VABench uses Qwen3-ASR-1.7B on the speech subset. The authors say so themselves, and quoting one as the other is the easiest mistake to make with this release.

• WeMM-Video-4B, joint audio-visual — Video-MME with audio 83.9 against Qwen3-Omni-30B-A3B-Instruct's 77.3 and MiniCPM-o-4.5-9B's 69.2; Video-MME v2 48.1 against 36.3 and 28.1; Daily-Omni 82.5 against 70.7 and 80.2, all in Thinking mode.

• WeMM-Video-4B, where it loses — MMAU 75.2 against Qwen3.5-Omni-Flash's 80.4 and MMAR 67.3 against 74.0, with one win on MMAU-Pro at 68.1 against 64.8. Video-Holmes 58.8 against MiniCPM-o-4.5-9B's 64.3, and FutureOmni 50.9 against 55.1 and 56.1.

Two caveats travel with the WeMM column. The paper does not state the frame count, pixel budget or context length used for its video evaluations, which matters because the shipped recipe uses sixteen frames where the processor defaults to 128. And the mode changes between tables: Thinking for video, Instruct for audio, Thinking again for image and text. Comparing a WeMM audio row against a rival video row is not a comparison.

Wall-clock, the number nobody quotes

Kandinsky's repository publishes the figure that decides whether an open video model is usable on your hardware, and it is not flattering. These are the non-distilled base model's working times for a single five-second clip after warmup, with weight loading and MP4 encoding excluded, Full HD meaning the super-resolution pass is running too.

• Full HD, Pro — 402 seconds on an H100, 765 on an RTX PRO 6000, 854 on an RTX 5090, 1,106 on an A100 80GB, 1,247 on an RTX 4090, 3,530 on an RTX 5060 Ti.

• Full HD, Lite — 284 seconds on an H100, 387 on an RTX PRO 6000, 406 on an RTX 5090, 664 on an A100, 578 on a 4090, 1,774 on a 5060 Ti.

• SD, Pro — 356 seconds on an H100 against 936 on a 4090, where the consumer-card penalty is smallest.

Peak allocated memory on a Pro SD run reaches 72.8 GiB, past every consumer card, which is why the repository ships block offloading — two transformer blocks resident, the rest streamed from host memory — with presets for 32, 24 and 16 GB. The 32 and 24 GB presets are identical because above 24 GB the headroom buys no speed. The buried caveat is that on the 16 GB preset the text encoder is quantised to NF4, and the paper is explicit that this is the only preset change that alters the clip itself, at around 12 percent of pixels differing by one brightness level at roughly 57 dB PSNR. If you are reproducing someone else's result on a 16 GB card and it does not match, check that first.

WeMM-Video-4B has no equivalent table, and that absence is itself informative. T​encent validated that the checkpoint loads and generates. A model whose selling point is video, described only as "it generated," leaves the entire cost side of a deployment decision to you.

Wiring a text layer around both

If you run either of these, the surrounding stack needs the same three things, and none of them is the model. Prompt expansion turns a shot description into the long, medium or short caption Kandinsky was trained on, or into the question WeMM should answer. Transcription and captioning feed an image-to-video pass or prepare a clip for description. Review summaries decide which of four generations to keep, which is the call you pay for on every candidate rather than every survivor.

Those are ordinary text completions, and they are the part of this stack OrcaRouter covers: one OpenAI-compatible key across more than two hundred models, provider list prices passed through with 0% markup so a vendor price change lands the same day, and automatic failover so a caption stage that dies does not take a five-minute render with it. Neither model is on the platform — the catalogue has no Kandinsky entry and no WeMM entry, and no routing layer turns a weight download into a hosted endpoint. What the router removes is the integration work around the two models you do self-host, plus a way to hold a second vendor behind the first without writing a second client.

Two ceilings, one decision

Ask which cap you can work inside. If you need footage with sound and you can cut in five-second units, Kandinsky 6.0 Video is the more complete release on every axis that is not raw quality: MIT licence, four deployment surfaces, an H100 number you can plan a budget around, and a super-resolution pass that makes Full HD a real setting rather than a claim. Its constraint is duration, and duration is easy to design around.

If you need to know what is in footage you already have — what happens on screen and what is audible over it — WeMM-Video-4B is the only one of the two that can answer, and its constraint is fidelity: sixteen sampled frames and eight thousand tokens are what the shipped recipe gives you, and the long-video benchmark rows were produced under conditions the paper does not describe.

What neither release provides is the thing that would make this comparison cheap: a number from somebody who did not write the paper. Both are open, both are permissively licensed, and both are invisible to every leaderboard that would settle the argument. Until one of them is measured independently, the ceiling in each repository is the most reliable fact either vendor has published.