
WeMM-Video-9B vs Alibaba Wan 2.7: A Bill and a Checkpoint
- 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 · 356 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 · 147 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 · 92 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 · 81 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 117 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
Start with the invoice. Alibaba Wan 2.7 has a per-minute rate on two Artificial Analysis video boards and a documented duration range, which makes it one of the few video models you can cost out before you write a line of code. WeMM-Video-9B has no rate at all. Tencent's WeChat Vision group published it on 17 November 2026 as nineteen and a half gigabytes of Apache-2.0 weights — five BF16 shards, a model card, and an arXiv preprint numbered 2611.26409 — and then said nothing. No launch post, no pricing page, no hosted endpoint, no board submission. Three days later the repositories still show zero downloads between them. So the honest comparison is not Wan 2.7 against WeMM-Video-9B on quality; it is a metered service against an unpriced artifact, and the two numbers you would want to put side by side do not exist for the same reason on both sides.
Everything below separates three things: what the Wan 2.7 documentation and the two leaderboards state, what the WeMM-Video-9B repository and paper claim, and what nobody outside either lab has reproduced. That third bucket is where most of the WeMM-Video-9B material sits, and pretending otherwise would be the fastest way to make this piece useless.
Wan 2.7 is billed in minutes
Alibaba's rate card for Wan 2.7 is unusually legible for a video model. The suite is a four-mode product — text-to-video, image-to-video, reference-to-video and video editing — and the pricing that surfaces on the boards is a straight per-minute figure rather than a credit system you have to reverse-engineer.
• AA-Video-Editing v1.1 — Wan 2.7 sits 9th at Elo 1,074 ±8 over 22,754 samples, a release month of April 2026, and $16.90 per minute. The interval is the tightest on the board outside the top three, which is what twenty-two thousand votes buys you.
• AA-Video-T2V v2.1 — the same family appears as Wan2.7-260612 at 20th on Elo 1,035 ±13 over 6,965 samples, June 2026, $9.00 per minute. Same vendor, different task, a different price, and an Elo that does not convert into the editing score.
• Wan 2.7 is not on the image-to-video board at all. Wan 3.0 is — 12th at 1,109 — and that absence is worth reading carefully. It does not mean Wan 2.7 cannot animate a still; it means the arena has not ranked it there. The Alibaba suite does advertise first-and-last keyframe control, up to five subject images and five reference clips, and 720p through 1080p output with native audio. Capability the board does not score is still capability.
One contextual note that matters for budgeting: Wan 2.7 is the previous generation. Wan 3.0 now leads both the text-to-video board (1,176) and the editing board (1,197), and it is the cheaper of the two on text-to-video at $12.00 per minute against Wan 2.7's $9.00 editing rate. If you are choosing an Alibaba model today rather than comparing one, the decision has already moved on.
What the 9B checkpoint actually is
WeMM-Video-9B is not a video generator, and that single sentence does most of the work in this comparison. It is a video understanding model: you hand it frames plus the original soundtrack, and it returns text. Description, transcription, temporal reasoning, question answering over footage you already hold.
The repository is unusually specific about what is inside it, which is the one place a quiet release beats a loud one. The Hugging Face record lists 10,488,603,384 parameters across 1,305 tensors, BF16, 19.54 GiB split over five shards. The split is itemised: a 7.94B language backbone including a 1.02B token embedding, an untied 1.02B output head, a 1.08B Whisper-style audio tower with 32 layers at width 1,280 behind a two-layer Q-Former, and a 0.46B vision tower of 27 blocks at hidden width 1,152 with patch size 16 and a 2×2 spatial merge. The language stack is 32 layers at hidden 4,096, 16 attention heads against 4 key-value heads at head dimension 256, with hybrid linear attention — one full-attention layer per three linear ones — and 262,144 position embeddings.
What that buys you in practice is a wide input and a narrow output. Video, audio, images and text go in. Text comes out, in English or Chinese. There is no frame synthesis anywhere in the architecture, and no amount of prompting will produce one.


Two clocks that do not meet
The comparison that actually decides anything is about time and money, and both sides have a cap.
• Wan 2.7's cap is duration. Two to fifteen seconds per generation, 720p to 1080p, billed per second. A fifteen-second 1080p clip is a known quantity before you press go.
• WeMM-Video-9B's cap is context. The processor defaults to 2 frames per second with a minimum of 4 frames, a maximum of 128 frames and a 100,663,296-pixel total budget. The shipped vLLM recipe overrides all of that down to 2 fps and 16 frames at 262,144 pixels — roughly eight seconds of coverage sampled from the midpoint of the clip. Every serving command in the repository also passes an 8,192-token context length, and at ten audio tokens per second that is about thirteen minutes of audio if you spend the budget on sound alone.
• The mismatch. Wan 2.7 produces fifteen seconds of footage. WeMM-Video-9B, as shipped, looks at eight. If your pipeline generates a clip and then asks a model what happened in it, the describer is sampling half the output. That is not a flaw in either product; it is a configuration you have to change deliberately, and the repository's own example does not change it for you.
The board neither of them is on
Here is the structural problem with this matchup, stated plainly. Every Artificial Analysis video board ranks generated footage. Text-to-video, image-to-video, video editing — all three take a model that outputs pixels and rank it by pairwise human preference. WeMM-Video-9B outputs tokens.
It is not that WeMM-Video-9B scored badly. It cannot be entered. Wan 2.7 has two rows, both with intervals tight enough to argue about; WeMM-Video-9B has none and will not have any until somebody builds a video-understanding arena. Which means the only quality evidence for the subject of this article is the authors' own table.
That table is not nothing. In thinking mode on joint audio-visual tasks, the 9B checkpoint beats Qwen3-Omni-30B-A3B-Instruct on 35 of 37 shared figures — Video-MME with audio 85.1 against 77.3, Daily-Omni 83.6 against 70.7, SciVideoBench 64.2 against 49.9, OmniReasoningBench 43.9 against 24.1 — and loses exactly one, FutureOmni at 53.7 against 55.1. On silent video against its own Qwen3.5-9B backbone it gains on F-16-NBA (88.7 vs 75.8) and TimeLens, and loses on MLVU, LVBench, VideoVista, VSI-Bench and MVBench. The paper says it reports "official results where available and our own evaluations otherwise," which is the correct disclosure and also the reason none of it is independent.

Where OrcaRouter fits
Neither model is served through OrcaRouter, and this piece is not going to pretend otherwise. Wan 2.7 is reachable through Alibaba's own API and several third-party platforms; WeMM-Video-9B is reachable by downloading it.
What does belong here is the billing side, because the reason teams end up comparing a generator with a describer is that both get wired into the same pipeline. OrcaRouter passes provider list price through at 0% markup, which means a vendor rate cut lands in your invoice the day it lands in theirs rather than whenever a reseller re-prices. If Wan 2.7's $16.90 editing rate becomes $12.00 the way Wan 3.0's did, you see it on the next bill. For a workload where one leg is metered per second and the other leg is a checkpoint you host yourself, having the metered half report at cost is the difference between a budget you can forecast and one you cannot.
Which one to budget for
If the deliverable is footage, Wan 2.7 is the only one of the two that can produce it, and the decision reduces to whether $16.90 a minute for editing or $9.00 a minute for text-to-video is competitive against the rest of the board. At 9th and 20th respectively, it is mid-field on both.
If the deliverable is a statement about footage you already hold, WeMM-Video-9B is the only one of the two that can produce it, and the decision reduces to whether you can carry a patched vLLM or SGLang install, a 19.54 GiB checkpoint, and an eight-second sampling window to get a 9B-class describer with genuinely strong audio-visual numbers.
The trap is reading $16.90 and $0.00 as the same quantity. One is a metered rate for a service you can call this afternoon. The other is the absence of a service, and the price of that absence is paid in integration work and in the fact that nobody outside Tencent has told you whether the numbers hold.
