2026-11-10 · VideoFreeTier

Seedance 2.0 open source: the most-liked Hugging Face repo is 4.7 KB and its README says the weights are not in it

In the channels we read on 10 November 2026, nobody is publishing Seedance 2.0 weights. The most-liked repository returned by the Hugging Face search for seedance holds exactly two files, 1,519 and 3,190 bytes, and its own README says the full model is hosted on a platform of its own rather than in the repository. Fourteen different accounts carry that README byte for byte. That is a finding about these channels, not a claim that no weights exist anywhere: we did not check every possible repository name, every hosting platform, or private repositories, and we did not download or run any weight file.

Everything below was read on 10 November 2026 through public APIs without registering, subscribing, buying credits, or running a generation. Four channels are in play and none speaks for the others: the Hugging Face Hub model search and blob listings, GitHub repository and tree endpoints, BytePlus ModelArk documentation, and one public Hacker News thread. A repository name containing the word seedance is not evidence that a model is inside it; a like count is not a download count, and a star count is not evidence that weights exist. Where a page prints nothing, this page says so instead of filling the gap. The dated observations behind the model pages live on the data page.

The top result is two files, and neither is a model

GET /api/models?search=seedance&limit=100&full=true on Hugging Face returned 32 repositories at 09:49 GMT+8 on 10 November 2026, the same count as the 08:51 capture. Sorted by likes, the first is Semub/Seedance-2-AI-Video-Generator-free: created 2026-06-19, zero downloads, and 101 likes at the later read against 100 at the earlier one. Its ?blobs=true listing contains two files and nothing else.

That total is 4.7 KB when a kilobyte is taken as 1,000 bytes, our arithmetic on the two published sizes. There is no .safetensors, no .pt, no .bin, no checkpoint directory, and no model card with a base-model field. One measurement detail matters for anyone repeating this: the full=true response lists 161 sibling entries across the 32 repositories and not one of them carries a size key, so a reader who judges emptiness from full=true alone would see no file sizes at all rather than small ones. Requesting ?blobs=true for each repository is what returns the bytes. This page pulled that endpoint for all 32.

The likes number is not a popularity measurement of a model and is not a download number. Every one of the 32 repositories in this result set reports zero downloads in its own ?blobs=true response, including the four that do hold large files. Zero downloads is what the API field says; it is not a claim about what people did outside it.

Fourteen accounts, one README

The 3,190-byte README of the top result has SHA-256 938b6332263f48c9a9ebe55fb887f1dc212fceaedb1a58824a12eae30d689e55. Pulling resolve/main/README.md from every repository in the result set shows the identical hash under 14 accounts, with like counts ranging from 0 to 101 and the same 4,709-byte two-file layout. The other copies include VikaskumarDas/Seedance-2-AI-Video-Generator, silenox/Seedance-2-AI-Video-Generator, RTX2o/Seedance-2-AI-Video-Generator, Arfoot/Seedance-2-AI-Video-Generator, Walee09/Seedance-2-AI-Video-Generator-free, Zane91/Seedance-2-AI-Video-Generator-free, svm-kids/Seedance-2-AI-Video-Generator-free, Axdix/Seedance-2-AI-Video-Generator, JRay0421/Seedance-2-AI-Video-Generator-free, janakbhavan/Seedance-2-AI-Video-Generator-free, Ai-PixVerse67/Seedance-2-AI-Video-Generator-free, aoo-thae-2026/Seedance-2-AI-Video-Generator-free and sleepydoug1992/Seedance-2-AI-Video-Generator-free. Three further pairs share their own hashes, and 28 of the 32 repositories have a README at all, across 12 distinct hashes.

The shared text is explicit about where the model is. It reads Due to the massive computational resources required to run the Dual-branch DiT architecture, the full Seedance 2.0 model is currently hosted exclusively on our official platform. and, immediately after, You do not need to download heavy weights, configure environments, or rent expensive GPUs. Apart from two status badges on img.shields.io, every link in that README points to https://seedance2.plus, including the Experience Seedance 2.0 here call to action and the contact section. So the repository that ranks first on likes describes itself as a pointer to a browser site, not as a weight release.

Two boundaries belong here. This README is a claim written by whoever uploaded these repositories; it is not a statement from ByteDance or BytePlus, and this page did not verify who runs seedance2.plus or whether it is affiliated with the model's maker. The README also prints claims about output resolution and usable-output rate that this page does not adopt, because they are marketing copy on a third-party upload with no measurement behind them. What is verifiable is the file list, the byte sizes, the hash, and the text as printed.

The four repositories that do hold weight-like files

Across all 32 blob listings, exactly four repositories contain files with the extensions .safetensors, .pt, .bin, .ckpt, .pth, .onnx or .gguf. None of the four is a Seedance 2.0 video model.

The three yang29 repositories are named for seedance15, not for 2.0, and their contents are the parts of a training run: a model checkpoint, a discriminator, an optimizer state, per-rank files and LoRA adapters, including lora/adapter_model.safetensors at 700,000,440 bytes and checkpoint-step-60/discriminator.pt at 147,873,012 bytes. We did not read their training configuration or establish what those checkpoints were trained on. The fourth is the only repository whose name contains seedance-2.0 and that actually carries a .safetensors file, and it is the clearest case of a name not describing the contents: the README's model-type field says LoRA, its frontmatter has no base_model key, and its own description names minimax-h3 as the fine-tune source. The same page also links out to a service that runs Seedance 2.5 through an API. Nothing in the repository states a base model of Seedance 2.0, and we did not load the file in ComfyUI to see which model it works with.

Two more repositories are worth naming because they are bigger than all of those and still are not video-model weights. xiwenyoumu/pi05-beat-drum-seedance100-v1 and xiwenyoumu/pi05-weigh-apple-seedance100-v1 hold 24 files each at about 12.44 GB, but their own model cards describe vision-language-action policies fine-tuned from pi05_base for a robot arm, with library_name: openpi. Their READMEs say the training video was rendered with a video generation model named as Seedance, and one carries a warning that it is a diagnostic checkpoint rather than a validated controller. Those are research artifacts that used generated video as training data, not a release of the generator's weights. A large file size therefore does not answer the question either.

The 8,109-star GitHub repository is a prompt pipeline

On GitHub the wording of the query changes the answer, so this page records both. Restricting to repository names with q=seedance in:name&sort=stars returns 2,062 repositories and puts Emily2040/seedance-2.0 first at 8,109 stars. Without the in:name restriction, q=seedance&sort=stars returns 63,927 repositories whose first result is lsdefine/GenericAgent at 14,402 stars, with the seedance-named repository in third place. Saying it is the top result for a plain seedance search would be wrong; it is the top result when the search is limited to names.

The repository itself is MIT licensed, 33,342 KB, created 2026-02-25, and describes itself as Comprehensive production pipeline for quad-modal AI filmmaking with Seedance 2.0. A recursive git/trees/HEAD read returned 397 entries with truncated false and no file ending in .safetensors, .pt, .bin, .ckpt, .pth, .onnx or .gguf. Its top-level directories are .github, agents, assets, data, docs, evals, examples, references, schemas, scripts, skills, tests and validation. That is a repository of prompts, schemas, skills and tests built around a hosted model, which is a different thing from a repository containing that model. A star count says people found the pipeline useful; it says nothing about whether weights were published.

A separate query, q=seedance org:bytedance, returned total_count: 0 on 10 November 2026. That is a result about a GitHub organisation and a search term, not proof that the company has no public code anywhere, and it does not speak to other hosts or to repositories under other organisation names.

What the official documentation publishes instead

BytePlus ModelArk serves seedance-2-0.md as a plain Markdown document; the copy read on 10 November 2026 was 77,050 bytes. Its model table prints exactly three identifiers:

A case-insensitive scan of that document returns zero lines containing weight, zero containing safetensors, zero containing open source and zero containing huggingface. It does contain the word download five times, and every one is about something else: an example attachment image to unzip, the uv tool, Python 3.12, the URL of a finished video, and a SKILL.md file. Its eight GitHub mentions are all Go import paths under github.com/volcengine/ark-runtime-go/. The code samples call the model through an API with a model ID string. So the official channel publishes an API contract, and this document contains no weight download and no statement about open-sourcing.

That is as far as the official text goes. We did not find a BytePlus page that says the model will never be released as weights, and we did not find one that says it will be. Whether the vendor intends a future weight release is 未公布 / 本站未实测. An official API document is one channel; its silence about weights is a fact about that document, not a property of the model's architecture or licensing. An earlier guide on this site, Seedance 2.5 API endpoint, made the same distinction on a different version: a documented endpoint is not a downloadable model.

The Hacker News quote is about renting GPUs, not downloading a model

One widely quoted comment in the 2026-06-16 Hacker News thread Running local models is good now is easy to misread. Comment 48561450 by the account echelon, read through the public HN Algolia API on 10 November 2026, says For instance, ComfyUI is an abomination that can't do half of what Nano Banana and Seedance 2.0 can do. And you have to sit around and wait 10x longer for single results. and then I can rent an H200 for $3.50 an hour. That's INSANELY cheap. The same comment continues I do not understand this split between hosted APIs and rinky-dink local RTX models.

Read as a whole, the comment uses Seedance 2.0 as a cloud benchmark for what local hardware and ComfyUI cannot match, and argues for renting data-centre GPUs instead. It is not a report of having downloaded Seedance 2.0 weights, and it gives no location for any weight file. The rental price is the commenter's own figure, not a price this site verified, and it is not a model price. Quoting this comment as evidence that weights are available locally would reverse what it says.

What this page does not settle

No weight file was downloaded, no checkpoint loaded, no ComfyUI workflow run, no generation attempted, and no account created on any of the platforms named here. So this page does not establish whether the 626 MB LoRA can be loaded against any video model, what the three yang29 checkpoints were trained on, whether seedance2.plus has any relationship with the model's maker, or whether the vendor plans a weight release. Those are 未公布 / 本站未实测 where the pages and APIs are silent.

The scope is a defined set of channels at a defined time: the 32 repositories returned by one Hugging Face search query on 10 November 2026, four GitHub API calls, one official Markdown document, and one public comment thread. We did not enumerate repositories whose names omit the word seedance, other model hosts, private repositories, or anything published after that date, and we make no claim that no Seedance 2.0 weights exist anywhere. The phrase weight leak appears in autocomplete suggestions as a question people type; this page found no evidence of a leak and makes no such claim. Suggestion lists only tell us which phrasings people use, and this page quotes no ranking, volume, difficulty, bid, or traffic figures.

Every number above belongs to a named channel: the byte sizes, hashes, like counts and zero-download fields to the Hugging Face Hub API, the stars and tree contents to the GitHub REST API, the three model IDs to BytePlus ModelArk documentation, and the quoted sentence to one Hacker News commenter. None of them is a property of the model weights, and a statement about one channel is not a statement about the others. The dated observations behind the model pages live on the data page.

Related on this site: MiniMax H3 free — a case where weights really are published, with conditions attached. Seedance 2.5 API endpoint — what an official API contract looks like when there is no download. Seedance 2.5 API free — a free-token headline whose table does not cover 2.5. The observed figures behind the model pages live in the data page.