No. As of October 13, 2026, Seedance 2.5 is not open source. ByteDance has not published its weights, so you cannot download it or run it locally. The official ways to use it are the BytePlus ModelArk and Volcengine Ark APIs and ByteDance's own apps such as Dreamina. Third-party platforms, including SeedVideo AI, run it in the browser. For local generation, use an open-weight model such as Wan 2.2.
That answer matters if you are planning a pipeline, a budget, or a privacy review. Below you will find what ByteDance actually publishes, how Seedance 2.5 compares with open-weight video models on license and hardware, the three legitimate ways to use it, and how to spot "Seedance" repositories that are not what they claim to be.
Last verified: 2026-10-13. Licenses, repositories and access routes change, so check the linked sources before you commit.
Image prompt: A quiet home video studio at night, a compact desktop workstation with a glass side panel and one glowing graphics card on a walnut desk, a wide monitor showing a full-screen cinematic frame of a coastal cliff at sunrise, a ceramic mug and a notebook, warm lamp light mixed with cool monitor glow, shallow depth of field, 35mm editorial photography, 16:9.
TL;DR
- Seedance 2.5 is a closed, hosted model. There are no official weights on ByteDance Seed's Hugging Face or GitHub accounts.
- You can use it through the official BytePlus ModelArk or Volcengine Ark API, through Dreamina, or through third-party platforms such as SeedVideo AI.
- "Run Seedance locally" is not possible today. Repositories with Seedance in the name are prompt collections, skills or API wrappers.
- If you need weights on your own GPU, look at Wan 2.2 (Apache 2.0), HunyuanVideo 1.5 or LTX-2.5, and read each license first.
Quick Answer
| Question | Answer as of 2026-10-13 |
|---|---|
| Is Seedance 2.5 open source? | No. It is a proprietary ByteDance Seed model |
| Are Seedance 2.5 weights available? | No. No official checkpoint has been published |
| Can I run Seedance locally? | No. Every route runs on ByteDance or partner servers |
| Is there an official Seedance GitHub repo? | No official model repository exists |
| How do I get API access? | BytePlus ModelArk (international) or Volcengine Ark (mainland China) |
| What is the closest open-weight option? | Wan 2.2, HunyuanVideo 1.5 or LTX-2.5, depending on license and GPU |
If you just want to make a video, the Seedance 2.5 model page on SeedVideo AI lets you generate in the browser without any setup.
Seedance 2.5 vs open-weight video models
This table compares access and licensing. Visual quality depends on the prompt, the source material and settings, so judge it with your own side-by-side tests.
| Model | Weights published | License | Official GPU guidance | Commercial use | Audio |
|---|---|---|---|---|---|
| Seedance 2.5 | No | Proprietary, hosted only | Not applicable | Under the terms of the platform you use | Generates audio and video together |
| Wan 2.2 | Yes (Hugging Face, ModelScope) | Apache 2.0 | TI2V-5B: 24 GB GPU such as an RTX 4090; A14B models: 80 GB GPU in the official commands | Allowed under Apache 2.0 | No audio in text or image-to-video; a separate S2V model follows supplied speech |
| HunyuanVideo 1.5 | Yes (Hugging Face) | Tencent Hunyuan Community License | 8.3B parameters; 14 GB minimum with offloading | Allowed in the licensed territory, which excludes the EU, UK and South Korea | Not in the base model |
| LTX-2.5 | Yes (Hugging Face) | LTX-2.x Community License | No single minimum; depends on pipeline, resolution, quantization and offloading | Free below US$10 million annual revenue; larger companies need a paid license | Generates synchronized audio and video |
| Wan 3.0 | No checkpoint files yet | Repository lists Apache 2.0 | Hosted API only for now | Under Alibaba Cloud's API terms | Generates audio |
Seedance 2.5. Best for: 30-second clips with audio, reference-heavy work and teams that do not want to run servers. Not for: offline inference, fine-tuning, or anyone who needs to own the weights.
Wan 2.2. Best for: local generation under a permissive license, plus a large community of LoRAs and ComfyUI nodes. Not for: one-click 30-second clips with built-in sound.
HunyuanVideo 1.5. Best for: the lowest official memory floor of the three open models. Not for: companies operating in the EU, UK or South Korea, where the license does not apply.
LTX-2.5. Best for: local audio-and-video generation and high-resolution pipelines. Not for: companies above the revenue threshold that cannot sign a commercial agreement.
For a deeper look at the Wan side, see why Wan 3.0 shipped without open weights and the older Seedance 2.5 vs Wan 2.2 comparison.
What ByteDance actually publishes
Seedance comes from ByteDance Seed, the company's model research team. Seed does release some open models: its Hugging Face account lists 59 models, including Seed-OSS, Seed-Coder and the SeedVR video restoration models. None of them is Seedance. Its GitHub organization has 63 repositories, and none contains Seedance weights or inference code.
The official Seedance 2.5 page tells the same story. It offers three actions: Get API, Try now and Compare now. The Seed model directory links Seedance 2.5 to the BytePlus Playground and to Dreamina. There is no download link, model card or license file for local use.
ByteDance Seed's official Seedance 2.5 page. Box 1 marks the access options: Get API, Try now and Compare now.
Seedance 2.5 is described there as an audio-video joint generation model built for 30-second storytelling, with reference control and editing. Those features come with hosted access. You send a request; ByteDance or a partner runs the model and returns the file.
Three legitimate ways to use Seedance 2.5
1. Official API and ByteDance apps
Developers can call Seedance 2.5 through BytePlus ModelArk outside mainland China or through Volcengine Ark inside it. Both use an asynchronous task API: you submit a job, keep the task ID and poll for the finished video. Activation, billing and rate limits depend on the account. Our Seedance 2.5 API guide covers endpoints, request fields, pricing and limits.
Creators who do not code can use Dreamina, the app linked from ByteDance Seed's model directory.
Best for: production integrations that need an official contract. Not for: quick tests without a cloud account and a funded balance.
2. Third-party platforms, including SeedVideo AI
Some platforms connect to Seedance 2.5 through API providers and wrap it in a browser workspace. SeedVideo AI at seeddance.io is one of them. It is an independent platform, not a ByteDance product, and its account, credits and pricing are separate from Dreamina's. The Seedance vs Dreamina explainer goes through how those names relate.
The SeedVideo AI workspace with Seedance 2.5 selected: 1 is the prompt box (up to 10,000 characters), 2 is the settings row for ratio, resolution, duration and sound, and 3 is the model selector.
To try it, open the AI text-to-video generator, keep Seedance 2.5 selected, write a prompt and press Generate. Clips run from 4 to 30 seconds.
Best for: creators who want results today without API keys. Not for: anyone who needs the model inside their own network.
3. Open-weight alternatives on your own hardware
If the requirement is "no data leaves our machines" or "we need to fine-tune", Seedance 2.5 is out. Pick an open-weight model from the table above and plan for the hardware. Wan 2.2's TI2V-5B is the realistic starting point on a single 24 GB card. Its official notes say a 5-second 720p clip takes under 9 minutes on one consumer GPU without extra optimization. Larger models need datacenter cards or heavy offloading.
Budget for engineering time as well. Installing drivers, matching library versions and building a queue all cost hours that a hosted service absorbs.
Best for: private pipelines, research and custom fine-tunes. Not for: teams that want Seedance-level 30-second clips with sound from a single prompt.
Common misconceptions about Seedance and open source
"There is a Seedance GitHub repo with the model"
Search GitHub for "seedance" and the most-starred results are prompt libraries, prompting skills for coding agents, storyboard tools and API clients. Some are useful. None contains ByteDance weights. A repository that claims to host the "full Seedance 2.5 model" and asks you to run an installer deserves suspicion.
"Hugging Face has Seedance weights"
Hugging Face does show models with Seedance in the name. The ones we checked were uploaded by individual accounts, and most had zero downloads. None came from ByteDance. The official publisher would be the ByteDance-Seed account, and it lists no Seedance model.
"A site called Seedance is the official one"
Many sites use Seedance or similar spellings in their domain. Most, including seeddance.io, are independent services that access the model through an API. That is legitimate, but you should know who you are paying. ByteDance's own entry points are seed.bytedance.com, BytePlus, Volcengine and Dreamina.
"Seedance 2.0 was open, so 2.5 will be too"
No Seedance version has been released as open weights, including 1.x and 2.0. ByteDance could change course in the future, but nothing published so far points that way. Our Seedance models category tracks version changes as they happen.
Try Seedance 2.5 in a browser
The quickest way to judge whether you need local control is to see what the hosted model does with your own ideas. These three prompts work as written in the text-to-video workspace.
Video prompt: A courier in a bright yellow raincoat rides a bicycle through a narrow city street at night after rain, wet cobblestones reflecting warm shop lights and magenta neon, steam rising from a street-food stall, low tracking shot beside the bike, teal and amber grade, light film grain.
A courier in a bright yellow raincoat rides a bicycle through a narrow city street at night after rain. Wet cobblestones reflect warm shop lights and magenta neon. Low tracking shot beside the bike, teal and amber color grade, light film grain, sound of tires on wet stone.
Try this prompt in SeedVideo AI
Slow aerial push-in over a rugged coastal cliff at sunrise. Waves break white against dark rocks, sea mist drifts through golden light, seabirds glide past the camera. Natural colors, steady drone movement, one continuous shot with the sound of surf and wind.
Try this prompt in SeedVideo AI
A ceramicist shapes a tall vase on a spinning wheel in a quiet sunlit studio. Close-up on wet clay rising between her hands, then a slow pull back to reveal shelves of finished pots. Soft window light, calm documentary feel, gentle ambient studio sound.
Try this prompt in SeedVideo AI
If one of these looks like what you need, the hosted route is probably enough. If you keep running into the limits, such as no fine-tuning or the need for offline runs, that is the signal to test an open-weight model. To start from a still frame instead, use the AI image-to-video generator to animate a source image.
FAQ
Is Seedance 2.5 open source?
No. As of October 13, 2026, ByteDance has not published Seedance 2.5 weights, training code or an open license. It is available only as a hosted model through official APIs, ByteDance apps and third-party platforms.
Can I download Seedance 2.5 weights?
No official download exists. Files claiming to be Seedance weights on file-sharing sites, Hugging Face or GitHub are not from ByteDance. Avoid running installers from unknown sources.
Can I run Seedance locally on my GPU?
No. Seedance runs only on ByteDance or partner servers. For local generation, Wan 2.2 TI2V-5B runs on a 24 GB card, and HunyuanVideo 1.5 lists a 14 GB minimum with offloading.
Is there an official Seedance GitHub repository?
No. ByteDance Seed's GitHub organization has no Seedance repository. Popular "seedance" repos are community prompt libraries, prompting skills and API clients.
How do I get Seedance 2.5 API access?
Create a BytePlus ModelArk account (international) or a Volcengine Ark account (mainland China), activate the model and call the video generation task endpoint. Third-party providers also resell access with their own pricing.
What is the best open-source alternative to Seedance 2.5?
It depends on the constraint. Wan 2.2 has the most permissive license (Apache 2.0). HunyuanVideo 1.5 has a low memory floor but excludes the EU, UK and South Korea. LTX-2.5 generates audio but needs a paid license above US$10 million revenue. None matches Seedance 2.5's 30-second, audio-and-reference workflow out of the box.
Start with the hosted model, then decide
For most creators, the useful test is whether the hosted version already does the job. Open the Seedance 2.5 page on SeedVideo AI, run a five-second test with one of the prompts above, and move to an open-weight model only if you need offline runs or custom training.
