ByteDance (Doubao/Seed)

Thesis

ByteDance's Seed team is a full-stack frontier lab, not a single-model shop: it spans foundation LLMs, speech, vision, world models, robotics, agents, and AI infrastructure, with labs across China, Singapore, and the U.S. P23. It is large and commercially backed — roughly 2,000 employees led by former Google DeepMind research VP Wu Yonghui W5 — and it pairs unusually deep open-source systems releases (VeOmni, Triton-distributed, ByteCheckpoint, ShadowKV) P14P15P13P8 with consumer-first product shipping (Doubao, Jimeng, Seedance, SeedRealtime) W3W1. Its flagship general models stay closed-weight W5W1, while its public GitHub/Hugging Face footprint exposes a highly legible data/eval/infra lane: benchmarks (EdgeBench, EvaLearn, DAComp) P5P24E48, data-centric training (Seed-Coder, VideoWorld) P19P11, and inference/training systems P8P14P15. For a data-business operator, the actionable read is that ByteDance Seed is publicly signaling demand in data synthesis, evaluation, and distributed training/inference infrastructure, even as its commercial model strategy stays consumer-first and largely closed.

Signal desks

Hiring

• Research Scientist, LLM Foundation Models — Reasoning: the role spans reasoning/planning across data acquisition, model evaluation, pretraining, SFT, reward modeling, and RL, and explicitly calls for synthesizing large-scale high-quality (multi-modal) data via rewriting, augmentation, and generation, plus system-2 decoding such as MCTS and A* W2. Direct data-business lane: data acquisition, synthetic multimodal data, and model evaluation are named workstreams W2.
• Talent bet: ByteDance began placing heavier bets on AI talent recruitment in mid-2024 W4.
• Scale: the Seed model team is ~2,000 employees across China and abroad W5, with labs in China, Singapore, and the U.S. and research spanning LLMs, speech, vision, world models, AI infra, and next-generation interfaces P23.

Forks

• ByteDance-Seed/triton is a fork of triton-lang/triton E60, consistent with the lab's compiler/kernel investment in Triton-distributed P15. Beyond this, the fork desk is thin in this pack: decoupleQ is built on top of OPTQ P7 and Triton-distributed is based on OpenAI Triton P15, but these are derivative builds rather than captured fork events. No additional fork evidence cited.

Releases

• Seed-OSS-36B-Instruct (Apache-2.0, 36.2B params) is the highest-traction model release in the pack: 41,092 downloads and 505 likes E1, with Base and Base-woSyn variants E3E5.
• Stable-DiffCoder-8B (Instruct and Base, MIT, 8.25B params) shipped Jan 2026; Instruct has 142 likes on only 424 downloads E2E12.
• A steady stream of Apache-2.0 model cards shipped across agents (M3-Agent-Control, M3-Agent-Memorization) E6E17, provers (BFS-Prover-V2-7B/32B) E22E32, long-context finetunes (AHN-Mamba2/DN/GDN for Qwen-2.5) E16E27E34E35E45E55E56E57E58, biology (PAR, cryofm-v1/v2, ConfRover) E11E36E37E38E43, and editing (VINCIE-7B) E23.
• Infrastructure ships via release tags too: VeOmni v0.1.11 (May 2026) E46.

Talking

• SeedRealtime (Aug 2026): a native audio-visual full-duplex LLM that watches, listens, and speaks in one model; live in the Doubao app, with no technical report, parameter count, open weights, or API endpoint published W1W6.
• Seedance 2.5 (Jul 2026): video generation with 30-second clips, multi-turn extension, up to 30 images / 10 videos / 10 audio inputs; rolling out to Jimeng AI and Doubao Pro, with a Volcano Engine Ark API expected W3.
• Commercialization framing: Seedance is ByteDance's current media-generation wedge, while coding is framed as the missing piece where Doubao 2.1 must earn a seat W4.
• Scale framing: Doubao is China's most popular AI app at 300M+ monthly users, and ByteDance keeps weights of its most powerful general models confidential, with Seed2.1 (Jun 2026) positioned for reasoning, coding, multimodal, and agentic tasks W5.

Shipping

• Consumer/agentic: SeedRealtime shipped inside Doubao with no open artifacts W1; Seedance 2.5 shipped to Jimeng AI and Doubao Pro with an Ark API pending W3.
• Open-weight language: Seed-OSS-36B-Instruct/Base (Apache-2.0) E1E3; Stable-DiffCoder-8B-Instruct/Base (MIT) E2E12; Seed-Coder 8B family (MIT) P19; Seed-Thinking-v1.5 (20B active / 200B total MoE) P16.
• Open-weight multimodal: Bagel unified multimodal model (Apache-2.0) P18E4; Seed1.5-VL (532M vision encoder + 20B active MoE) P26.
• Systems: VeOmni with a released tag cadence (v0.1.11) E46, Triton-distributed P15, ByteCheckpoint P13, ShadowKV P8, FlexPrefill P9, SDP4Bit P10, decoupleQ P7.
• Science/biology: PAR protein checkpoints (400M/60M) P3P4; ByteCRN companion site P1.

Research themes

• Distributed training and inference systems: VeOmni (model-centric distributed recipe zoo, RL trainer) P14, Triton-distributed (distributed compiler) P15, ByteCheckpoint (checkpointing for LFMs) P13, ShadowKV and FlexPrefill (long-context inference) P8P9, SDP4Bit (4-bit training communication) P10, decoupleQ (2-bit PTQ) P7, StragglerAnalysis (training observability) P21.
• Agents and evaluation: Agent-R (reflective self-training) P12, M3-Agent ecosystem E6E17E30, EdgeBench (134 real-world tasks, 38,000+ hours of agent interaction, scaling laws) P5E14, EvaLearn (learning capability/efficiency) P24, DAComp (data agents across the data lifecycle) E48.
• Multimodal and vision: Bagel P18, SAIL (single-transformer vision-language, no pretrained vision encoder) P17, Seed1.5-VL P26, SeedVR/SeedVR2 (video restoration) P27, VINCIE (in-context editing) P28, ByteMorph (non-rigid image editing benchmark) P25, DeepFlow P22.
• World models and video: VideoWorld / VideoWorld 2 (learning from unlabeled video) P11E44, TraceAnything (4D trajectory fields) E47.
• Robotics: Chain-of-Action (trajectory autoregressive manipulation, NeurIPS 2025) P20, UAM (dual-stream VLA to avoid forgetting) P6E19.
• Test-time training: Modular-TTT (TTT as composable modules) P2, In-Place-TTT E53.
• Reasoning and code: Seed-Thinking-v1.5 P16, Seed-Coder (model-curated code data) P19, Seed-Prover / BFS-Prover (Lean) E31E22E32.
• Biology/chemistry: PAR (ICML 2026 Oral protein generation) P4, ByteCRN (chemical reaction networks) P1, cryofm E36E37.

Hiring & scaling

• The clearest single hiring artifact is the LLM Foundation Models — Reasoning research scientist role, which couples reasoning with data acquisition, model evaluation, pretraining/SFT/RLHF, synthetic multimodal data generation, and MCTS/A* decoding W2.
• Recruitment is a stated strategic lever: ByteDance began placing heavier bets on AI talent recruitment in mid-2024 W4, and the Seed team now numbers ~2,000 people across China and abroad, led by Wu Yonghui W5.
• Geographic footprint per Seed's own org page: China, Singapore, and the U.S. P23. No location-specific job postings beyond W2 are present in this pack, so hub-level hiring signals are thin beyond the stated multi-country footprint P23W2.

Data-business implications

• Data / synthetic data demand: the reasoning role explicitly tasks synthesizing large-scale high-quality (multi-modal) data via rewriting, augmentation, and generation across pretraining, SFT, and RLHF W2. Seed-Coder shows a model-centric data pipeline curating GitHub, commits, and code-related web data with minimal human effort P19. VideoWorld pursues label-free learning from unlabeled video P11E44, and SDP4Bit uses EleutherAI's the_pile_deduplicated as its training baseline P10. These collectively point to demand for scalable data acquisition, filtering, and synthesis tooling W2P19P11P10.
• Evals: ByteDance Seed is a benchmark producer, not just a consumer — EdgeBench (134 tasks, 12+ hour per-task agent runs, scaling-law framing) P5E14, EvaLearn (NeurIPS 2025, 5/5/5/5) P24E49, DAComp (full data-intelligence lifecycle) E48, ByteMorph (image editing) P25, plus internal benchmarks BeyondAIME and Codeforces from Seed-Thinking-v1.5 P16 and 38-of-60 benchmark claims from Seed1.5-VL P26E28. This signals a receptive buyer/evaluator for eval datasets, rubrics, and harnesses.
• Infrastructure / tooling: trainer-free and omni-model-native training (VeOmni, including an RL trainer) P14, a distributed compiler (Triton-distributed) P15, unified checkpointing (ByteCheckpoint) P13, training-communication quantization (SDP4Bit) P10, inference KV-cache and sparse-attention systems (ShadowKV, FlexPrefill) P8P9, and quantization (decoupleQ) P7 — a broad, actively released infra surface E46.
• Deployment: Seedance 2.5 is rolling out to Jimeng AI and Doubao Pro with a Volcano Engine Ark API expected W3; Seed1.5-VL was released on Volcano Engine P26; SeedRealtime is live in Doubao but has no third-party endpoint yet, which is itself a deployment-gap signal for integrators W1.
• Safety: no cited evidence in this pack on safety, alignment, or red-teaming posture. The absence is notable given the reasoning/RLHF role description W2; treat safety as an evidence gap.
• Product / GTM: Doubao is positioned at 300M+ monthly users and is China's most popular AI app W5; Seedance currently sells more than LLMs, with coding named as the missing commercialization piece for Doubao 2.1 W4. This frames immediate GTM focus on consumer media generation plus a coding/agent push W3W4W5.

Traction highlights

• Seed-OSS-36B-Instruct: 41,092 downloads, 505 likes E1.
• Bagel: 6,000+ stars (page capture) P18; event capture 6,149 stars, 7 HN points/5 comments E4.
• VeOmni: 2,003 stars (page) P14; event capture 2,156 stars E24.
• Triton-distributed: 1,457 stars P15 (event 1,517) E29.
• Seed1.5-VL: 1,580 stars P26 (event 1,583) E28.
• SeedVR: 1,222 stars P27 (event 1,317) E33.
• VideoWorld: 790 stars P11 (event 795) E44.
• ShadowKV: 306 stars P8 (event 311) E50; Seed-Thinking-v1.5 812 stars P16; m3-agent 1,442 stars E30; EdgeBench 68 stars page / 426 event P5E14; EvaLearn 431 stars P24E49.
• Note: star counts differ between page captures (dated 2026-06-11) and event captures (various dates); both are cited where available.