
WeMM-Video-35B-A3B vs Alibaba Wan 2.7: 22,754 Votes vs Zero
- 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
There is exactly one number in this comparison that a third party produced, and it belongs to Alibaba Wan 2.7. On Artificial Analysis' video-editing board the model sits ninth at Elo 1,074 with a confidence interval of ±8, computed from 22,754 pairwise human votes — the smallest interval anywhere on that board. WeMM-Video-35B-A3B has nothing comparable. Its Hugging Face repository was created on 17 November 2026 holding a single .gitattributes file, its 67.39 GiB of BF16 tensors only arrived at 03:57 UTC on 20 November, and as of this writing the page records zero downloads and three likes. The two models are not ranked against each other anywhere, and no amount of careful prose can manufacture that ranking.
What can be done is to separate the three kinds of claim in play: what 22,754 people actually judged, what Tencent's own paper asserts, and what neither side has shown. Most coverage of a fresh checkpoint blurs the second into the first. This piece keeps them apart, because the difference between a measured Elo and a vendor table is the whole decision.
The only third-party number here
Wan 2.7's editing row is unusually well supported. Twenty-two thousand votes is not a leaderboard sample you can wave away as noise; the ±8 interval means the ninth-place position is stable against everything below it, and the release month attached to the row is April 2026, so the votes have had seven months to accumulate.
The same family appears twice more on the boards, and the three rows do not agree with each other, which is itself informative:
• AA-Video-Editing v1.1 — 9th, Elo 1,074 ±8, 22,754 samples, April 2026, $16.90 per minute.
• AA-Video-T2V v2.1 — listed as Wan2.7-260612, 20th, Elo 1,035 ±13, 6,965 samples, June 2026, $9.00 per minute. Same vendor, different task, a third of the sample count, and a price that is 47% lower than the editing rate.
• AA-Video-I2V v2.1 — no row. Wan 3.0 is on that board at twelfth with 1,109; Wan 2.7 is not. The suite advertises image-to-video capability, so this is an arena gap rather than a product gap, but it means the model has never been ranked on the task most people would reach for first.

Read the three together and the honest summary is that Wan 2.7 is a measured mid-field video model with one genuinely strong evidence base (editing), one thinner one (text-to-video), and one blank cell. That is still vastly more external evidence than the other side of this comparison has.
What Alibaba is selling, and at what rate
Wan 2.7 is a four-mode product: text-to-video, image-to-video, reference-to-video and video editing. The documented controls are first-and-last keyframe conditioning, up to five subject images plus five reference clips, 720p through 1080p output, native audio, and clips from two to fifteen seconds. Billing is per second at the board rate, which makes a fifteen-second 1080p generation a cost you can compute before you press go rather than a credit balance you discover afterwards.
The awkward fact for anyone budgeting on Wan 2.7 today is that Alibaba has already superseded it. Wan 3.0 leads the text-to-video board at 1,176 and the editing board at 1,197, and it does so at $12.00 per minute — cheaper than Wan 2.7's $9.00 text-to-video rate only when you compare like for like, and cheaper than the $16.90 editing rate outright. Wan 2.7's April release month and its seventh-month-old votes describe a model that is now the previous generation inside its own vendor's lineup. That matters for this comparison because it changes what you are actually weighing: not the best Alibaba can do against the best Tencent can do, but a mature, measured, one-generation-old service against a checkpoint whose weights are hours old.
The checkpoint that landed this morning
The WeMM-Video-35B-A3B release is legible in a way most quiet releases are not, because Hugging Face timestamps the whole sequence. The repository appeared at 08:53:50 UTC on 17 November with nothing in it but a gitattributes stub. The arXiv report, 2611.26409, was submitted on 19 November at 16:58 UTC and has already been revised once, at 06:15 UTC on 20 November. The BF16 weights commit is dated 2026-11-20T03:57:36Z. The GitHub repository, Tencent/WeMM-Video, was created on 17 November at 02:29 UTC, carries nine stars, zero forks and zero published releases, and was still having its README edited at 06:34 UTC this morning.
So the paper preceded the tensors by roughly eleven hours, and the model card is still being rewritten as this is written. That ordering is the story. There has been no launch post, no pricing page, no hosted endpoint, no press release and no announcement account — nothing that a vendor normally does when it wants a model noticed.
What the repository does contain is specific:
• Size — 36,177,564,536 parameters, BF16, spread over seventeen safetensors shards totalling 67.39 GiB. The "35B-A3B" label counts something different from the file size; the gap is described further down.
• The FP8 build did come back. When the paper went up, the FP8 Block128 repository the card links answered 401 — the same answer the Hub gives for a name that has never existed. It resolves now: 36.09 GiB in one safetensors file, committed at 05:53:01 UTC, two hours after the BF16 upload, with 33.6B parameters in E4M3 and 2.56B left in BF16. Whatever was happening at 16:58 UTC on 19 November, the release was not finished.
• Three more quantised variants — AWQ W4A16 at 22.72 GiB across 52 per-layer shards, NVFP4 W4A16 at 24.11 GiB in one file, and a GGUF Q4_K_M bundle at 25.46 GiB split into a 22.61 GiB language file, a 0.84 GiB projector and a 1.99 GiB audio encoder. Two of those three describe themselves as unvalidated in their own manifests. The NVFP4 one claims only that the weights were re-read on CPU, which is a checksum and not an evaluation.
• Runtimes — nothing here runs on a stock engine. The repository ships source overlays with SHA-256 manifests for exactly four upstream releases: vLLM 0.18.0 and 0.31.0, SGLang 0.5.16 and 0.5.19. Every published serving command passes an 8,192-token context, tensor parallelism across two GPUs, and a media pipeline capped at 16 frames and 262,144 pixels. A search of the vLLM issue tracker finds no upstream request for this architecture.

What Tencent measured, and against whom
Every quality figure attached to WeMM-Video-35B-A3B comes from the team that trained it. The paper says as much — "we report official results where available and our own evaluations otherwise" — and the one number on the card whose conditions are given completely is not even a video result: MMAR1000, audio only, decoded at 16 kHz, greedy decoding, two H20s. The rest of the card records the checkpoint loading. Against 22,754 votes, that is a different category of evidence and should be read as one.
The headline table is a different matter, and it is large. Across 30 single-valued benchmarks, the checkpoint in thinking mode finishes ahead of Qwen3.5-Omni-Flash on 28 and behind on 2; ahead of Seed2.0-Mini-0428 on 26 and behind on 4; and behind Gemini-3.0-Flash on 21 of 30, which Tencent's own summary describes as narrowing a gap rather than closing it. The two rows that carry the most weight for anyone wiring up a describer are the audio-visual ones — 86.4 on Video-MME with audio against 79.3, and an eighteen-point lead on the harder v2 split at 54.9 against 36.6 — because a model that only reads frames cannot post either number.
The two losses point the other way and are worth repeating: JointAVBench at 65.9 behind Qwen's 69.0, and MUSIC-AVQA at 79.5 behind Qwen's 82.1. Both are audio-heavy. A bigger checkpoint does not win everywhere, and the places it does not win are the places where Qwen's audio stack has had longer to mature.
None of that has been reproduced. Neither Artificial Analysis nor LMArena lists a WeMM model of any size, and the three likes on the repository are not a peer review. If you are choosing between these two models to do a job, you are comparing 22,754 votes against a table signed by the authors.
Both take video, only one returns it
The structural difference is easy to state and hard to overstate. Wan 2.7 accepts text, images, reference clips and video, and returns video. WeMM-Video-35B-A3B accepts text, images, audio and video with its soundtrack, and returns text. Wan 2.7's editing mode comes closest to the other side — it takes footage in and gives footage back — but the output is still pixels, and there is no setting that makes it describe what it edited.
The caps on each side bound the pipeline you can build:
• Wan 2.7's cap is duration. Two to fifteen seconds per generation, 720p to 1080p, billed per second. A fifteen-second clip at the editing rate is about $4.23 before any retries.
• WeMM-Video-35B-A3B's cap is context and sampling. The processor defaults to two frames per second, a minimum of four frames, a maximum of sixteen and a 262,144-pixel total budget — roughly eight seconds of coverage at the shipped recipe. The serving commands cap context at 8,192 tokens and allow four images, two videos and two audio clips per request. At ten audio tokens per second, that context is about thirteen minutes of audio if you spend it on sound alone.
• The mismatch. Wan 2.7 can produce fifteen seconds of footage; the shipped WeMM recipe looks at eight. If your pipeline renders a clip and then asks a model what happened in it, the describer as configured samples about half the output. That is a configuration you have to change deliberately, not a limit of the architecture — the processor's own default allows 128 frames.
Two providers, one model, one key
Neither of these models is served through OrcaRouter. The catalogue lists no WeMM checkpoint of any size, and it lists no video generator at all — of the 183 model IDs available today, 51 accept video as input and not one returns video. Wan 2.7 is reachable through Alibaba's own API; WeMM-Video-35B-A3B is reachable by downloading 67 GiB and applying a source patch.
What the routing layer does offer is a way to stop paying for a single provider's bad afternoon. Qwen3.6-35B-A3B — the dense-ish backbone Tencent built this MoE on, and the model you would reach for if you want "watch the clip and reason about it" without the 67 GiB download — is in the catalogue twice: once at $0.248 per million input and $1.485 per million output with a 262,144-token context, and once, through a second provider serving the same weights, at $0.31 and $4.21. That is the same model at two prices, and the routing DSL lets you express the preference once — cheapest first, fail over on a 5xx or a timeout — instead of hard-coding a base URL into your client and hoping.
Automatic failover is the part that matters for video work specifically, because video requests are long. A 40-minute audio track at ten tokens per second consumes most of an 8,192-token budget, which means a single job occupies a connection for a long time, and a provider incident mid-request is expensive rather than annoying. Passing provider list price through at 0% markup matters for the other half of this comparison: when Alibaba cuts Wan 3.0 to $12.00 or a successor undercuts it again, a pass-through bill shows the cut on the next invoice rather than whenever a reseller re-prices.

Which number you are actually choosing
If the deliverable is footage, Wan 2.7 is the only one of the two that produces it, and the decision reduces to whether $16.90 a minute for editing or $9.00 a minute for text-to-video is defensible against a board where the current leader charges $12.00 and beats it on both tasks. Ninth and twentieth are respectable mid-field positions. They are not a reason to pick the previous generation.
If the deliverable is a statement about footage you already hold, WeMM-Video-35B-A3B is the only one of the two that produces it, and the decision reduces to whether you can carry a patched inference engine, a 67.39 GiB download or a 36.09 GiB FP8 one, two GPUs of tensor parallelism, and an eight-second sampling window — to get a 36B-class describer whose audio-visual numbers are strong and entirely first-party.
The trap is treating "Elo 1,074 over 22,754 votes" and "86.4 on Video-MME with audio" as commensurable. One is a crowd measurement of a service anyone can call. The other is a vendor table about a checkpoint that had no weights in it until this morning. Both are worth reading. Only one of them is evidence.
