
WeMM-Video-2B vs Seedance 2.5 Flash: The QC Gate Nobody Prices
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 184 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 535 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 262 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 94 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 291 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 99 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 132 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 84 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 122 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 104 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 · 89 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 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 · 216 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 108 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
Seven cents buys one second of Seedance 2.5 Flash at 480p, or fifteen cents at 720p, on the rate card ByteDance published through BytePlus on 13 November 2026. Nobody publishes what it costs to decide whether the clip that came back is usable. WeMM-Video-2B — Tencent's 3.28-billion-parameter omni-modal checkpoint, uploaded on 17 November 2026 with no announcement and, as of 20 November, zero downloads — is the first openly licensed model in this family small enough to run as the gate rather than the generator. Put together, they describe a pipeline where the expensive part has a price and the part that decides whether you paid it twice does not.
That is the whole argument, and it is not a product pitch for either model. Seedance 2.5 Flash is a video generator with a rate card and no capability table. WeMM-Video-2B is a video understanding model with a repository and no independent evaluation. Neither is on a hosted endpoint we can point you at. What follows is what each one actually is, why the price-per-second framing that dominates this category is incomplete, and what a gate in the loop would have to look like before it was worth wiring up.
What seven cents a second does not cover
Start with the arithmetic, because it is the only part of this comparison that is settled. Seedance 2.5 Flash bills $0.07 per second at 480p and $0.15 at 720p with no input video, against $0.103 and $0.231 for Seedance 2.5 — roughly two-thirds the price at both tiers, derived from a token rate of $7.00 per unit against the parent's $10.70. A five-second clip is $0.34 at 480p. A thousand of them is $340.
Those numbers are correct and they are also the least interesting thing about the model, because they price a generation and not a decision. Every studio that runs a video tier eventually discovers that the bill is not the number of clips it generated — it is the number it generated to get one it would ship. If Flash takes four attempts to land a usable take, its real cost is 28 cents per accepted second, and it beats its own headline by nothing at all. If the parent model lands in two, the parent is cheaper per accepted second despite costing 47% more per attempt. The rate card cannot tell you which of those is true. Only a measurement can.
That measurement is what an understanding model is for — and this is where the comparison stops being an apples-to-apples "vs" and becomes a description of two adjacent jobs.
One generates, one judges, and neither does both

Seedance 2.5 Flash takes a prompt, an optional first frame and reference assets, and returns video. WeMM-Video-2B takes video frames plus the original soundtrack, and returns text. There is no overlap in capability and no way to substitute one for the other. The model identifier for the generator is dreamina-seedance-2-5-flash-261115, and the date stamp in it — 15 November 2026 — is the API window BytePlus advertised, two days after the model surfaced on a distribution partner.
• What goes in — a prompt and reference images for Flash; up to four images, two videos and two audio items per request for WeMM-Video-2B.
• What comes out — mp4 video with synchronised audio for Flash; a text response for WeMM-Video-2B, in Chinese or English.
• Resolution — 480p and 720p on Flash, with no 1080p tier at all; not applicable to WeMM-Video-2B, which has no video output.
• Duration — five seconds is the worked example on the rate card; WeMM-Video-2B samples at 2 fps with a 4-frame floor and a 128-frame ceiling, so a five-second clip is ten frames by default.
• Weights — Flash is a closed endpoint with nothing to download; WeMM-Video-2B is Apache 2.0 and downloadable, but you supply the GPUs.
• Independent evaluation — none for either. No third-party leaderboard we could find on 20 November carries a Seedance 2.5 Flash entry, and the WeMM-Video paper's headline results belong to the 35B-A3B and 9B, not the 2B.
Read that list twice and the shape of the thing becomes clear. This is not a choice between two video models. It is a choice between paying for generation and paying for judgement, and most teams are already paying for the first one without having costed the second.
What Tencent actually built
The 2B is the compact member of a four-size family — 2B, 4B, 9B dense and a 35B-A3B mixture of experts — released together as seventeen Hugging Face repositories between 08:53 and 08:59 UTC on 17 November, with the arXiv paper following on 19 November and the repository README still being edited on the morning of the 20th. The licence is Apache 2.0 and the architecture is a three-tower stack under one `WeMMVideoForConditionalGeneration` wrapper: a 24-block vision encoder at 1,024 hidden with 16-pixel patches, a 32-layer Whisper audio encoder at 1,280 wide with 128 mel bins, and a 24-layer text model at 2,048 hidden, 8 attention heads and 2 key-value heads, on a 248,320-token vocabulary with a 262,144-position maximum.
The details that decide whether it is usable as a gate rather than a demo:
• Weight size — 6.12 GiB of BF16 safetensors across two shards, 6,567,256,208 bytes in the index, which is roughly 3.28 billion parameters once the encoders are counted. The name counts the language model only.
• Two modes, one checkpoint — `enable_thinking=False` returns a direct answer and `enable_thinking=True` returns a reasoning trace, selected by a chat-template flag rather than by downloading a second model.
• Serving — patched vLLM 0.18.0 or 0.31.0, patched SGLang 0.5.16 or 0.5.19, CUDA 12.9 images, and a native runtime path for the GGUF build. Tencent's own validation claim is that patched vLLM "loaded the checkpoint on L20 and H20 and completed generation checks for text, image, audio, and video with a soundtrack."
• Quantisation — FP8 Block128, AWQ W4A16, GGUF Q4_K_M and NVFP4 W4A16, each in its own repository. A 2B model in a 4-bit build is small enough to sit on the same box as the client that calls the generator.
• Input limits — four images, two videos and two audio items per request inside an 8,192-token context, enforced by the serving recipe rather than suggested by it.
What is missing matters as much. There is no 2B-specific accuracy number anywhere in the release — the family's claims about beating Qwen3.5-Omni-Flash and Seed2.0-Mini-0428 belong to the 35B-A3B, and the 9B's claim about surpassing Qwen3-Omni-30B-A3B-Instruct is the 9B's. Whether the compact model retains the audio-visual reasoning that makes the family worth writing about is, in public, untested. Zero downloads and two likes on the repository as of 20 November is a fair proxy for how much outside scrutiny it has had.
The capability table ByteDance has not published

Flash has the opposite problem. Its price is documented to the cent and its capability is not documented at all. BytePlus published a rate card, a model identifier and a resolution ceiling. It has not published, on any page we could load on 20 November, the maximum duration, the reference budget, whether audio is generated in the same pass, whether the editing and extension modes that define Seedance 2.5 carry over, or whether the mov output format is available.
The parent model answers all of those — four to thirty seconds, fifty reference assets split thirty images, ten videos and ten audio, joint audio-video generation, eleven languages, mp4 and mov at 24 fps. None of it is stated for Flash. That is not a reason to assume the worst; it is a reason not to assume anything. A tier that keeps thirty seconds and fifty references at two-thirds the price is a different product from one that keeps 720p and nothing else, and the difference is exactly the thing a QC gate would have to measure — except the gate cannot measure a spec that was never written down.
Which is the honest limitation of the whole idea. You can build the loop today. You cannot yet say what the loop is testing against, because one side of it ships a price list and the other ships a document.
What a gate in the loop would cost
Suppose you build it. The generator call goes to BytePlus or to a distribution partner on their key. The gate is a WeMM-Video-2B instance you host yourself, taking the returned clip and its soundtrack and answering one question — does this match the brief, and is the audio in sync. At 2 fps and ten frames for a five-second clip, that is a small inference job on a single GPU, and the 4-bit builds make it smaller still.
The economics are lopsided in a way that is worth stating plainly. A generation costs seven to fifteen cents per second and cannot be inspected before you pay. A gate pass costs a fraction of a cent of GPU time and can be run as many times as you like. If the gate rejects one in four clips, you have saved three-quarters of a regeneration you would otherwise have bought blind — and you have done it without a second vendor contract, because the gate is weights on your own hardware.
What you do not get is a number. Nobody has published an agreement rate between a WeMM-Video-2B judgement and a human editor's, and until someone does, the gate is a filter with an unknown false-positive rate. That is a real reason to run it in shadow mode for a week before you let it reject anything.
Where the routing layer earns its place

Neither model is on our catalogue, and both statements are worth making without hedging: we do not serve WeMM-Video-2B, and we do not serve Seedance 2.5 Flash. A pipeline built around either one is a pipeline built on somebody else's key.
What we do route is the surrounding work, and in this particular design that is not a consolation prize. Tencent Hy4 Flash is the cost-efficient tier of Tencent's Hunyuan flagship — text and image in, text out, a million-token context, up to 64,000 output tokens, at $0.0175 per million input tokens and $0.07025 per million output. That is the model that expands a one-line brief into the shot list, rewrites the prompt between attempts, and reads the gate's rejection reason to decide what to change. It is a text model doing text work, and at those rates the prompt-engineering loop around a video pipeline costs less than the video.
The structural point is the one that survives whichever vendor you pick. Provider list price passes through at 0% markup, so when ByteDance cuts a tier — the way it cut this one to two-thirds of the parent — the cut is live on our side the same day rather than at the next billing reset. Automatic failover across providers means the text half of the loop does not take the pipeline down when one endpoint degrades. And having 200-plus models behind one OpenAI-compatible key means the comparison you will inevitably run — this gate against a larger understanding model, this prompt-expander against a cheaper one — costs a parameter change and not an integration.
What to watch, and in what order
Three things would move this from an argument to a recommendation. First, the Flash capability table: duration, references, audio, and whether 1080p is permanently absent or merely late. Second, any independent evaluation of the 2B — a single public agreement number against human judgement would turn the gate from a design into a component. Third, the first published latency comparison for Flash, because "faster" is the entire premise of the name and 24 fps output at 720p for thirty seconds is a great deal of inference to accelerate.
Until then the summary is two prices and one missing number. Seedance 2.5 Flash is a real, cheap, narrowly specified video tier at two-thirds of its sibling's rate. WeMM-Video-2B is a complete, openly licensed understanding model that nobody has announced and almost nobody has downloaded. The seven cents is documented. The cost of knowing whether the seven cents was worth spending is not — and that gap, not the rate card, is the thing to budget for.
