Seedance 2.5: What ByteDance's Video Model Actually Is

Seedance 2.5 is ByteDance Seed's audio-video joint generation model, announced on 31 July 2026, and it is the current video model in ByteDance's Seedance line — its predecessor, the Seedance 2.0 series, is still on sale beside it rather than replaced by it. It generates video and audio in the same pass, accepts up to 50 reference assets in one request, and produces a coherent clip of up to 30 seconds that can then be extended in further rounds. And the answer to the question most people arrive with: the weights are not open, and ByteDance has published no plan to open them. There is no licence, no checkpoint and no repository for this model — the only thing the vendor ships is a hosted API and a set of consumer surfaces. That is the plainest fact about Seedance 2.5 and it is the one this page exists to state, because the searches that led you here are variations on "Seedance 2.5 open weight" and "Seedance 2.5 weights", and the honest answer is no.

This is a reference page rather than a launch report. The model is roughly eleven weeks old as I write, which is old enough that a good deal of what circulates about it — including on this blog — has drifted from what ByteDance actually documents today. Everything below is re-established from the vendor's own material: the ByteDance Seed model page and launch post, and the BytePlus ModelArk tutorials and pricing tables. Where the vendor publishes nothing, that is said rather than filled in.

What Seedance 2.5 is

ByteDance Seed's own description is compact: "a next-generation audio-video joint generation model, built for 30-second storytelling with precise reference control and powerful editing capabilities." The company's launch post frames the generation as a response to a shift in what people expect — "from merely generating a clip to completing a creative work" — and puts the emphasis on long-form storytelling, multimodal referencing and editing rather than on raw visual fidelity.

It sits on the architecture of Seedance 2.0, which ByteDance describes as a unified multimodal audio-video joint-generation design. Seedance 2.5 keeps that spine and pushes three axes at once. It is not a new family and it is not a different modality; it is the same line with the ceiling raised on duration, on reference volume, and on how precisely you can steer a shot after it exists.

What the vendor says about parameter count: nothing. Seedance 2.5's size, training data and compute are not published, and neither is any benchmark run by ByteDance. That absence is worth holding onto, because it means every quality claim about this model comes from one of two places — ByteDance's own marketing language, or a third-party human-preference board — and the two are not the same kind of evidence.

Are the Seedance 2.5 weights open?

No. Nothing about Seedance 2.5 is downloadable.

The vendor's material never mentions weights, checkpoints, a licence or an open-source release. The model page's calls to action are "Get API," "Try now" and "Compare now"; the footer links to models, publications and recruiting, not to a repository. BytePlus ModelArk serves the model as dreamina-seedance-2-5-260628, and its pricing page marks offline inference "Not supported yet" — there is no self-hosted path, only online inference against ByteDance's endpoint. A search of Hugging Face for the name returns no ByteDance-owned repository; the results are third-party pages that repackage API access under Seedance-flavoured titles, none of which are checkpoints.

Third-party classification agrees, and this is the more useful signal because it was not written by ByteDance. Artificial Analysis runs an explicit "Open weights" filter on its video leaderboards, and its own FAQ names the open-weights leaders on the text-to-video board: MiniMax H3 (768p) at Elo 1137, then LTX-2.5 Fast at 946 and LTX-2.5 Pro at 943. Dreamina Seedance 2.5, sitting at Elo 1143 and third overall on the same board, is not among them. The board's open-weights cohort and its Seedance 2.5 entry are disjoint sets.

It is worth being precise about what this does and does not mean. Closed weights do not make the model worse, and for a great many teams the question is academic: they were never going to run a video diffusion model on their own hardware. What closed weights do mean is that you cannot audit the model, cannot fine-tune it on your own footage, cannot pin a version and keep calling it after the vendor deprecates it, and cannot produce output without sending your prompts and reference assets to ByteDance. If any of those four constraints is disqualifying for your work, Seedance 2.5 is not your model, regardless of how the output looks. The open-weights options on that same board — MiniMax H3, LTX-2.5 — are where that search should end instead.

What it generates, at what length, at what resolution

BytePlus ModelArk's Seedance 2.5 tutorial is where the real specification lives, and it is considerably more specific than the launch post. The model decides its task type from the input assets and the prompt intent, and there are five recognised shapes:

• Text-to-video — a text prompt alone, with no special constraint on aspect ratio or duration.

• First-frame and first-last-frame video — one image as the opening frame, or two images as first and last, with the output aspect ratio locked to the first-frame image.

• Reference-to-video — at least one reference image, video or audio, with no duration constraint on the output.

• Video editing — a reference video plus an edit intent in the prompt, replacing a subject, adding or removing objects, or repainting part of the frame. Output keeps the source's aspect ratio and duration, the reference video must run 4–30 seconds, and the ratio must be set to adaptive with duration -1.

• Video extension — a reference video plus an extension intent, continuing the footage forward or backward. The launch post describes extending in further rounds while holding characters, environments and narrative pacing steady.

On the numbers the tutorial gives:

• Duration — 4 to 30 seconds in a single pass, or -1 to let the model pick a length inside that range. The Seedance 2.0 series tops out at 4–15 seconds, so this is a doubling, and ByteDance says it comes without multi-segment stitching.

• Resolution — 480p and 720p at 8-bit colour depth, and 1080p at 10-bit. That is the full list for Seedance 2.5. Not 4K. The 4K tier exists in the Seedance 2.0 series, not here, and this is one of the figures that has drifted: several third-party write-ups — and this blog's own earlier explainer — described Seedance 2.5 as a 4K model. BytePlus ModelArk's current capability table does not list 4K for dreamina-seedance-2-5-260628. Check the vendor table against your own delivery spec before you promise anyone 4K out of this model.

• Aspect ratios — 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 and adaptive.

• Container — mp4 or mov. The mov path uses H.264 with yuv444p chroma sampling and PCM audio, which ByteDance positions as the higher-fidelity option for editing and extension work.

• References — up to 50 assets in one request: 30 images, 10 videos and 10 audio clips. The 2.0 series allows 15 in total, split 9/3/3. Reference videos may total no more than 30 seconds of input, each must be 200 MB or less, and accepted input frame rates run 24–60 fps.

• Not published — the output frame rate. The tutorial specifies the frame-rate range it accepts on input, and specifies nothing about what it emits. If your pipeline assumes a fixed output fps, that is an assumption, not a documented value.

One further detail that has no counterpart in the 2.0 series: Draft mode. You generate a 480p preview to check scene structure, shot scheduling, subject motion and prompt alignment, then feed that draft task ID forward to render the final version at your target resolution, reusing the model, prompt, input assets, seed, audio settings, aspect ratio and duration. Only 480p drafts are permitted. Both steps are billed separately.

Audio is generated, not attached

Seedance 2.5 is an audio-video joint generation model, and the distinction is not marketing. The API carries a generate_audio flag, and the launch post's own examples are written as sound design: a "ding" discovery sting on a character's reaction, a crack as a grapefruit peel splits, a thump followed by half a second of silence, a voiceover reading a brand line over a white frame. Those are authored as part of the prompt that produces the clip, not layered on afterwards.

The reference side changed too. Seedance 2.0's audio references were conditional — an audio reference required an accompanying image or video. Seedance 2.5 accepts reference_audio on its own, and permits audio combined with image, with video, or all three together. Reference audio and video assets total no more than 30 seconds combined.

Watermarking is a parameter rather than a fixed behaviour: watermark is false by default, and setting it true stamps an AI-generated marker in the lower-right corner. Languages supported natively number eleven — Chinese, English, Spanish, Indonesian, Malay, Thai, Arabic, Portuguese, Vietnamese, Japanese and Korean.

What ByteDance claims, and what the independent boards measure

ByteDance publishes no benchmark for Seedance 2.5. Not on the model page, not in the launch post, not in the ModelArk tutorial. The company's claims are qualitative and are best read as a list of intentions: smoother and more consistent motion, more realistic visuals, stronger shot transitions and scene changes, better handling of object and skin texture, lighting and colour saturation, and a reduction in the uncontrolled subtitles and background music that make AI video look synthetic. The launch post also volunteers a limitation, which is unusual enough to quote: ByteDance says there is "still room for improvement, particularly regarding the physical plausibility of complex motions and the stability of scenes involving interactions among multiple subjects."

The independent picture comes from Artificial Analysis, which ranks video models by Elo from blind pairwise votes cast by a recruited human panel. On AA-Video-T2V v2.0, the board for text-to-video with audio, Dreamina Seedance 2.5 holds Elo 1143 with a 95% interval of 1134 to 1152, across 8,264 votes, placing it third. Ahead of it: Wan 3.0 at 1156 and Utopai X at 1149. Behind it: MiniMax H3 (768p) at 1137 and MiniMax H3 Max at 1130.

Three cautions on that number. The board's own name for the model is "Dreamina Seedance 2.5," and its listed release is July 2026 — this is the model in question, not its predecessor. The rating is a preference measure, not a capability score: it says people picked its clips more often in blind comparisons, which is a statement about output as judged by eye and ear, and nothing about controllability, API ergonomics, cost or reliability. And the intervals overlap heavily — 1143 against 1137 is not a ranking anyone should build a decision on.

Two further placements round this out. On AA-Video-T2V-Silent v2.0, the no-audio board, Seedance 2.5 sits at Elo 1105 in fourth, behind Wan 3.0, MiniMax H3 Max and MiniMax H3 (768p) — a consistent but unremarkable position. On AA-Video-I2V v1.0, the image-to-video board, it does not appear at all; the Seedance entries there are Dreamina Seedance 2.0 720p and Seedance 1.5 pro. Nothing here supports the claim that Seedance 2.5 leads its field. Nothing here contradicts it either — third overall on the audio-inclusive text-to-video board is a genuinely strong debut, and the board's own note that six models were added in the last thirty days says the ranking is still moving.

What it costs, and how you get to it

Access runs through two entirely different doors. On the consumer side, ByteDance's launch post names Jimeng AI and Doubao Pro — select Seedance 2.5 in the video generation menu — with the BytePlus playground at ai.byteplus.com and Dola listed as further entry points on the model page. On the developer side, the API is BytePlus ModelArk, model ID dreamina-seedance-2-5-260628, served from the Asia-Pacific endpoint. Before the model will run at all, BytePlus requires one of: an account balance above USD 30, a dedicated AI Savings Plan at the USD 30 tier or above, or a purchased Seedance 2.5 series resource pack. An API key alone is not enough, and that gate is worth knowing before you plan a spike.

Pricing is published, and it is metered by estimated token consumption rather than by the clip. The token rates, from BytePlus ModelArk's pricing page, in USD per million tokens:

• 480p and 720p output — 10.70 without video input, 6.40 with video input.

• 1080p output — 11.70 without video input, 7.00 with video input.

• Offline inference — not supported.

Token consumption scales with duration, resolution and frame rate, and BytePlus's own worked examples make the shape concrete. For a 5-second 16:9 clip with no video input, the vendor's estimate is USD 0.514 at 480p (0.103 per second), USD 1.156 at 720p (0.231 per second) and USD 2.843 at 1080p (0.569 per second). Once video goes in as an input, the same 5-second output runs USD 0.553–2.152 at 480p, USD 1.244–4.838 at 720p and USD 3.062–11.907 at 1080p, with the low end corresponding to a 2–4 second input and the high end to a 30-second one. Note that adding video input lowers the per-token rate while raising the total, because the input duration itself consumes tokens. Artificial Analysis normalises the same model to $34.12 per minute of 1080p at default settings, which is consistent with the vendor's own 1080p per-second figure once you multiply it out at 60 seconds.

For scale, the Seedance 2.0 series is cheaper across the board: 0.07 per second at 480p, 0.15 at 720p, 0.37 at 1080p and 0.78 at 4K on the same pricing page. Seedance 2.5's 1080p rate is roughly 1.5 times its predecessor's, and its 480p rate roughly 1.5 times as well. You are paying for duration, reference capacity and editing control, not for a cheaper meter.

One practical note on the meter, because it changes what you should measure: BytePlus charges only for successfully generated videos and does not bill a failed generation caused by content moderation. In an editing workflow, where output duration is inherited from the source clip, that matters more than in a prompt-driven one — a rejected edit costs you nothing but the retry, which is not true of every video API. Evaluate this model on cost per accepted clip, not cost per second of nominal output.

The routing question, honestly

Seedance 2.5 is not routed on OrcaRouter. A check of our live model list on 13 October 2026 returns no Seedance entry of any generation, so this page makes no claim about calling it through us, and you should treat any write-up that implies otherwise as out of date.

Where OrcaRouter does fit is the rest of the pipeline. A video generation model is rarely the only thing in the request path — there is prompt drafting, reference captioning, shot-list generation, moderation passes and the agentic orchestration around all of it, and that work does not belong on ByteDance's endpoint. OrcaRouter gives you one OpenAI-compatible endpoint across 200+ models with 0% markup, passing the provider's list price through unchanged, so a vendor price cut shows up on our side the same day it lands. Automatic failover means a prompt-drafting step keeps running when one provider has a bad afternoon, and the routing DSL lets you express which model handles which call as code rather than as a stack of if-statements. Bring your BytePlus key for the video generation and route everything around it through one place.

What is still open

The questions a reference page cannot settle, and the ones worth watching:

• Whether 4K arrives. The 2.0 series has a 4K tier and Seedance 2.5 does not, which is an odd inversion for a newer model and may be a staging decision rather than a ceiling. Watch the BytePlus capability table, not blog posts.

• Whether the weights ever open. ByteDance has said nothing either way. The Seedance line has never shipped a checkpoint, and the presence of genuinely open rivals on the same leaderboard — MiniMax H3 is the open-weights leader on the text-to-video board — is the commercial argument against changing that.

• Whether the Elo holds. Third place on AA-Video-T2V v2.0 with an interval overlapping the fourth-place model is a starting position, not a verdict, and 8,264 votes is a small base next to the boards' older entries.

• What the API surface settles into. Task-type inference from prompt intent is elegant in the examples and fragile in production — BytePlus documents two error codes, InvalidParameter.TaskTypeMismatch and InvalidParameter.TaskTypeConstraint, for exactly the case where you meant an edit and the model read a fresh generation. Setting omni_reference_task_type explicitly moves the validation earlier, and any real integration should.

• Whether ByteDance ships the physics. The company named complex motion plausibility and multi-subject interaction stability as its own weak points. Those are the two things that decide whether a 30-second clip is usable or merely impressive.