← Analysis archivereport.jsonevidence.jsonDossier →
ByteDance (Doubao/Seed)Frontier labgenerated Jun 8, 2026 · 3d

ByteDance (Doubao/Seed) analysis

27 cited sources99 signals on file

Thesis

ByteDance's Seed lab (Doubao/Seed) is shipping aggressively in the open, spanning a foundation-model line (Seed-OSS), specialized reasoning/coding/proving models, multimodal and 3D-vision systems, and the distributed-training and GPU-kernel infrastructure underneath them. The footprint reads as a full-stack research org: it open-sources both the models and the training/serving substrate (VeOmni, Triton-distributed) rather than weights alone. Breadth across modalities — language, vision-language, video, depth, code, theorem proving — is the defining signal.

Shipping

The flagship open release is [Seed-OSS-36B-Instruct](https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct) (~36B params), the lab's most-downloaded model at 39,280 30-day downloads and 499 likes, paired with [Seed-OSS-36B-Base](https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base) (1,852 downloads) and an ablation variant [Seed-OSS-36B-Base-woSyn](https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Base-woSyn) (116 downloads, 53 likes). The supporting repo [seed-oss](https://github.com/ByteDance-Seed/seed-oss) has 885 stars.

Beyond the base line, the lab ships narrowly specialized models:

  • [Stable-DiffCoder-8B-Instruct](https://huggingface.co/ByteDance-Seed/Stable-DiffCoder-8B-Instruct) (766 downloads, 137 likes) and its [Base](https://huggingface.co/ByteDance-Seed/Stable-DiffCoder-8B-Base) (389 downloads) — diffusion-style code generation.
  • [BFS-Prover-V2-7B](https://huggingface.co/ByteDance-Seed/BFS-Prover-V2-7B) (606 downloads) and [BFS-Prover-V2-32B](https://huggingface.co/ByteDance-Seed/BFS-Prover-V2-32B) (224 downloads) — automated theorem proving.
  • [cudaLLM-8B](https://huggingface.co/ByteDance-Seed/cudaLLM-8B) (178 downloads, 29 likes) — CUDA-kernel generation.
  • A family of AHN (Artificial Hippocampus Network) adapters for Qwen-2.5-Instruct at 3B/7B/14B in Mamba2, GDN, and DN variants (each ~50–66 downloads), e.g. [AHN-Mamba2-for-Qwen-2.5-Instruct-3B](https://huggingface.co/ByteDance-Seed/AHN-Mamba2-for-Qwen-2.5-Instruct-3B).
  • [M3-Agent-Control](https://huggingface.co/ByteDance-Seed/M3-Agent-Control) (~33B, 50 downloads / 50 likes), the control model behind the m3-agent line.

On GitHub the most-starred work skews multimodal and vision: [Bagel](https://github.com/ByteDance-Seed/Bagel) (5,994 stars), [Depth-Anything-3](https://github.com/ByteDance-Seed/Depth-Anything-3) (5,489), [Seed1.5-VL](https://github.com/ByteDance-Seed/Seed1.5-VL) (1,580), [m3-agent](https://github.com/ByteDance-Seed/m3-agent) (1,369), [SeedVR](https://github.com/ByteDance-Seed/SeedVR) (1,213), [VideoWorld](https://github.com/ByteDance-Seed/VideoWorld) (790), [Seed-Coder](https://github.com/ByteDance-Seed/Seed-Coder) (755), [Seed-Thinking-v1.5](https://github.com/ByteDance-Seed/Seed-Thinking-v1.5) (812), and [TraceAnything](https://github.com/ByteDance-Seed/TraceAnything) (540).

Infrastructure ships on its own cadence: the training framework [VeOmni](https://github.com/ByteDance-Seed/VeOmni) (1,991 stars) has the most active release stream — through [v0.1.11](https://github.com/ByteDance-Seed/VeOmni/releases/tag/v0.1.11) (plus v0.1.10 and several v0.1.9 alphas) — and [Triton-distributed](https://github.com/ByteDance-Seed/Triton-distributed) (1,455 stars) has cut tagged releases from experimental through [v0.0.2-rc](https://github.com/ByteDance-Seed/Triton-distributed/releases/tag/v0.0.2-rc). Smaller tooling releases include [JoltQC v0.1](https://github.com/ByteDance-Seed/JoltQC/releases/tag/v0.1) and a [bamboo_mixer manuscript](https://github.com/ByteDance-Seed/bamboo_mixer/releases/tag/manuscript) tag.

Research themes

No first-party writing captured yet. (Directions are nonetheless legible from the shipped artifacts: multimodal understanding/generation — Bagel, Seed1.5-VL; 3D and video perception — Depth-Anything-3, SeedVR, VideoWorld, TraceAnything; agents with memory — m3-agent / M3-Agent-Control and the AHN "Artificial Hippocampus" adapters; code and CUDA generation — Seed-Coder, Stable-DiffCoder, cudaLLM; theorem proving — BFS-Prover; and reasoning — Seed-Thinking-v1.5.)

Hiring & scaling

No careers data captured yet.

Traction highlights