
WeMM-Video-9B vs Kandinsky 6.0 Video: Both Open, Neither Measured
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 182 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 364 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 148 tok/s
- orcaNEWOrca: OrcaC2 1.02026-11-0737 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 93 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 135 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 82 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 118 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 110 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 66 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 91 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 42 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 211 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 109 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
This is a matchup where both models are open weights and both are free, which removes price from the comparison entirely and leaves two much harder questions: what each one will actually let you do, and whether anyone has checked. Kandinsky 6.0 Video, from Sber's Kandinsky Lab, was open-sourced on 6 October 2026 under MIT in two sizes — Pro at 29 billion parameters, Lite at 3 billion — with five-second clips, 44 kHz synchronised audio and lip-sync, and day-one support in Diffusers, ComfyUI, vLLM-Omni and SGLang. WeMM-Video-9B, from Tencent's WeChat Vision group, appeared six weeks later on 17 November 2026 under Apache 2.0 as 19.54 GiB of BF16 weights, unannounced, with no endpoint and no deployment story beyond a patched vLLM or SGLang overlay. Neither has an independent Elo anywhere. Neither will get one soon, for reasons that are worth separating.
What follows treats the two releases as what they are: a production-grade open model with an ecosystem, and a research release with a paper. Both descriptions are compliments and both are limitations, and the order you read them in will determine which one you should download.
Five seconds is the trained length
Kandinsky 6.0 Video produces five-second clips, and the important word is "trained," not "default." Five seconds at 121 frames is the length the model learned; asking for more is not a setting, it is extrapolation, and Sber does not present it as anything else. Inside that five seconds you get 44 kHz audio generated in the same pass with lip-sync alignment, which is a genuinely unusual capability at this licence and this size.
The deployment surface is the part that separates the two releases most sharply. On day one, Kandinsky 6.0 Video shipped a Diffusers bundle, a command-line entry point, two ComfyUI nodes and two Spaces. Then an outside contributor merged pull request 8537 into vLLM-Omni, registering the pipeline as Kandinsky6TI2VAPipeline and serving text-to-video-and-audio and image-to-video-and-audio from one DiT.
WeMM-Video-9B shipped a patched vLLM or SGLang overlay and a set of manifests. The only validation Tencent claims is that the patched runtime loaded the checkpoint on an H20 and completed a forward pass. Getting from there to a service is a day of toolchain work, and the repository is candid about it in the only way that matters: the status strings.
• FP8 Block128, 13.10 GiB — quantized_and_independently_reloaded_native_inference_pending, with native_inference_validated and native_low_precision_gemm_validated both false.
• AWQ W4A16, 12.86 GiB — quantized_unvalidated.
• NVFP4 W4A16, 13.07 GiB — cpu_NVFP4_weight_artifact_independently_reread, with native FP4 GEMM, PRO5000 inference and vLLM validation all false, plus a numeric attestation of mean relative RMSE 0.095 against a 0.35 limit.
• GGUF Q4_K_M, 13.28 GiB — exported_unvalidated, and it runs only on Tencent's own runtime, not llama.cpp or Ollama.
Four compressed builds, all within 0.42 GiB of each other, none of them validated for native inference. Compare that with Kandinsky 6.0 Video, where the MIT licence and four independent front ends mean the validation problem has already been distributed to people who are not Sber.

Nineteen and a half gigabytes, itemised
WeMM-Video-9B compensates for its thin deployment story with an unusually complete account of itself. The parameter count is 10,488,603,384 across 1,305 tensors. The split is published: 7.94B language backbone including a 1.02B token embedding, 1.02B untied output head, 1.08B audio tower — 32 Whisper-style layers at width 1,280 behind a two-layer Q-Former, ten audio tokens per second — and 0.46B vision tower at 27 blocks, hidden 1,152, patch 16, 2×2 spatial merge.
Only the dense FFN is quantised in the compressed builds: 96 language matrices, three per layer across 32 layers at 4,096 by 12,288, which is 4,831,838,208 elements out of 10.49B. Attention is restored to BF16 — 152 projection matrices plus 32 input norms, all with SHA-256 hashes — and 1,209 tensors stay bit-exact. Roughly 3.57B parameters, about 7.1 GB, of vision tower, audio tower, embedding and output head never enter the quantiser at all. That is why AWQ saves 34% against BF16 rather than the 70% the name suggests.
None of this tells you whether the model is any good. It tells you the release is auditable, which is a different and underrated virtue.
Neither has an independent score
Kandinsky 6.0 Video appears on none of the three Artificial Analysis video boards. Neither does WeMM-Video-9B. The two absences have different causes and the difference matters.
Kandinsky 6.0 Video has no arena submission, and the text-to-video board requires an API price before it will list a model — five seconds of open-weight output with no commercial rate has no row to occupy. WeMM-Video-9B is unrankable in a deeper sense: the boards score generated footage, and it generates text.
What each lab offers instead is its own evaluation, and both should be read with the same scepticism.
• Kandinsky 6.0 Video — vendor-reported VABench results plus a human evaluation run by the lab, roughly 200 pairwise comparisons per criterion. Also a reported 47% reduction in generated-speech word error rate after RL post-training. That last figure needs a flag: it was measured with Whisper-large-v3, while the VABench WER figure it is compared against uses Qwen3-ASR-1.7B. Two different recognisers, two different error rates, no valid subtraction. The improvement may well be real; the number as presented is not comparable.
• WeMM-Video-9B — the authors' own harness, disclosed as "official results where available and our own evaluations otherwise." In thinking mode on joint audio-visual tasks it beats Qwen3-Omni-30B-A3B-Instruct on 35 of 37 shared figures, including Video-MME with audio 85.1 against 77.3 and SciVideoBench 64.2 against 49.9, and loses one, FutureOmni at 53.7 against 55.1. On silent video against its Qwen3.5-9B backbone it gains on F-16-NBA 88.7 against 75.8 and TimeLens, and loses on MLVU, LVBench, VideoVista, VSI-Bench and MVBench.
Two open releases, two self-assessed evaluations, no third party. That is the honest state of the art for both.

Sixteen frames against five seconds
Where the two caps meet is more interesting than either one alone. WeMM-Video-9B ships a vLLM recipe that overrides its own processor defaults — 2 fps, minimum 4 frames, maximum 128, 100,663,296 pixels — down to 2 fps and 16 frames at 262,144 pixels, sampled around the midpoint of the clip. Sixteen frames at two per second is eight seconds of coverage, and every serving command also passes an 8,192-token context against 262,144 position embeddings in the architecture.
• Kandinsky 6.0 Video — five seconds in, five seconds out, 121 frames, 44 kHz audio. A hard trained length.
• WeMM-Video-9B — eight seconds of sampled input, text out, with the full audio track up to roughly thirteen minutes at ten tokens per second.
• The join — a five-second Kandinsky clip fits inside WeMM-Video-9B's eight-second window with three seconds to spare, which is the one configuration where routing a generated clip into the describer is clean.
• The cost of the join — Sber reports roughly seven minutes of H100 time per Full-HD clip, so a five-second generation is a seven-minute reservation before any describing happens on local hardware.
Routing by clip length
Neither model is served through OrcaRouter, and neither will be: both are open weights you host yourself. What a routing layer does for a pipeline like this is decide which leg runs where, and the routing DSL is the honest place to put the eight-second boundary. Express it as a policy — clips under eight seconds to the local WeMM-Video-9B describer, longer clips to a hosted omni model — and the boundary stops being a comment in a config file that nobody reads until a sixty-second clip comes back described in eight-second fragments.
For the hosted leg of that policy, one API covers 200-plus models with 0% markup on provider list price, so a pipeline that spans a self-hosted describer and a rented one is still a single bill and a single key. Provider rate cuts reach you the day they reach the provider, which matters when the hosted half of your workload is chosen on price.

Download both, or neither
If you want a working pipeline this week, Kandinsky 6.0 Video is the answer and it is not close. MIT licence, four deployment surfaces on day one, synchronised 44 kHz audio with lip-sync, and a five-second trained length you can plan around. You give up duration and independent validation.
If you want a 9B omni-modal describer you can audit, WeMM-Video-9B is the answer, and the audit is the point: full parameter accounting, itemised quantisation coverage, honest status strings on four unvalidated builds. You give up a serving contract, an ecosystem, and any evidence beyond the authors' own tables.
The mistake is reading "open weights" as a single property. One of these releases behaves like a product with a licence attached. The other behaves like a paper with a checkpoint attached, and which one you need depends entirely on whether you are trying to ship footage or understand it.
