OpenModelsHQ

ByteDance Seed

Beijing, China · ByteDance · China Desk

ByteDance Seed is based in Beijing, China and is part of ByteDance. We track 16 open-weight models from it, and the newest is Stable-DiffCoder-8B-Instruct, released 15 Jan 2026.

Models tracked
16
LLM 10 · Vision 5 · Image 1
Newest release
Stable-DiffCoder-8B-Instruct
15 Jan 2026
Country
China
Releases timeline
16 ByteDance Seed releases from 20 Jan 2025 to 15 Jan 2026. The largest is UI-TARS-72B-DPO at 73.4B parameters. Dot area follows total parameters; a ring means the size is not published.
Releases timeline: data
ModelCategoryReleasedParameters
Stable-DiffCoder-8B-InstructLLM15 Jan 20268.3B
BFS-Prover-V2-7BLLM6 Oct 20257.6B
Seed-OSS-36B-InstructLLM20 Aug 202536.2B
Seed-OSS-36B-BaseLLM20 Aug 202536.2B
VINCIE-3BImage2 Jul 2025Not published
BAGEL-7B-MoTVision19 May 202514.7B
Seed-Coder-8B-Reasoning-bf16LLM13 May 20258.3B
Seed-Coder-8B-ReasoningLLM27 Apr 20258.3B
Seed-Coder-8B-InstructLLM27 Apr 20258.3B
Seed-Coder-8B-BaseLLM27 Apr 20258.3B
UI-TARS-1.5-7BVision16 Apr 20258.3B
academic-ds-9BLLM9 Apr 20259.4B
BFS-Prover-V1-7BLLM23 Feb 20257.6B
UI-TARS-72B-DPOVision22 Jan 202573.4B
UI-TARS-7B-DPOVision22 Jan 20258.3B
UI-TARS-2B-SFTVision20 Jan 20252.4B

Origin

ByteDance, the Beijing company behind TikTok and Douyin, set up its Seed AI research team in 2023, according to Wikipedia. It launched the Doubao chatbot app the same year, in August. ByteDance itself dates from March 2012, when Zhang Yiming and Liang Rubo founded it.

The same article describes Seed’s main language model as proprietary. The open work is a separate stream, published on the ByteDance Seed page on Hugging Face and credited in model cards to the “ByteDance Seed Team”.

How it got here

Seed has no single open family. It ships different kinds of model.

UI-TARS, from 2025, operates graphical user interfaces and comes in 2B, 7B and 72B sizes. Seed-Coder, also from 2025, is a set of 8B code models with base, instruct and reasoning versions. BAGEL-7B-MoT, from the same year, handles multimodal understanding and generation in one model. Part of it comes from other labs’ open components: its card says it was fine-tuned from Qwen2.5-7B-Instruct and a SigLIP vision model, and that it uses the FLUX.1-schnell VAE.

Seed-OSS landed on 20 August 2025. The 36B language model comes as a base model in two versions, with and without synthetic data, and as an instruct model. Its card pitches it at long context, reasoning and agent tasks.

Two later releases are narrower. BFS-Prover-V2, from 2025, is a 7B theorem-proving model for Lean 4. Stable-DiffCoder, from 2026, is an 8B code model built on the Seed-Coder architecture with a diffusion-based training stage.

Licensing pattern

Seed sticks to standard permissive licences. UI-TARS, BAGEL and Seed-OSS carry Apache 2.0; Seed-Coder and Stable-DiffCoder carry MIT. We found no custom terms and no licence change between releases.

The BAGEL card states Apache 2.0 and names the parts it builds on: the Qwen2.5 model, the SigLIP encoder and the FLUX.1-schnell VAE.

Wikipedia describes Seed as proprietary and Seed-OSS as the open-source model. The hosted Doubao and Seed models are not open-weight releases, so we leave them out.

Release lineage

Newest first
  1. Stable-DiffCoder-8B-Instruct8.3B
  2. BFS-Prover-V2-7B7.6B
  3. Seed-OSS-36B-Instruct36.2B
  4. Seed-OSS-36B-Base36.2B
  5. VINCIE-3B
  6. BAGEL-7B-MoT14.7B
  7. Seed-Coder-8B-Reasoning-bf168.3B
  8. Seed-Coder-8B-Reasoning8.3B
  9. Seed-Coder-8B-Instruct8.3B
  10. Seed-Coder-8B-Base8.3B
  11. UI-TARS-1.5-7B8.3B
  12. academic-ds-9B9.4B

Models from ByteDance Seed

Compare in the table API pricing

Largest first. The last line on each tile is the memory of the smallest device that fits the model at 4-bit.