ByteDance

ByteDance Seed (with SUTD, Georgia Tech and M-A-P) shifts the unit of agent evaluation from task outputs to runnable infrastructure: since a model's c…

Tsinghua with ByteDance Seed, M-A-P, and TokenWave.AI loops the middle half of a MoE transformer's layers twice at matched FLOPs, parameters, and KV c…

ByteDance Seed (with SUTD, M-A-P and TokenWave.AI) reframes self-evolution evaluation: existing work hands the model human-specified tasks and metrics…

ByteDance Seed (with M-A-P and TokenWave.AI) asks whether an agent can actively test its own behavior, judge the resulting experience, and use it to i…

ByteDance Seed (with Nanjing University, M-A-P and TokenWave.AI) grounds an agent benchmark in market demand rather than researcher intuition: tasks a…

ByteDance Seed / Tsinghua study (work done at Seed; project lead Xintong Hao, corresponding author Jingzhao Zhang) of how many times to repeat scarce…

ByteDance's video-generation flagship: 30-second native-4K single-run generation ("one-take creation") with multi-turn extension, up to 50 multimodal…

Unified audio creation model: speech, sound effects, ambience, and music generated on one timeline in a single pass — multi-character dialogue,…

Visual reasoning without text chain-of-thought: UniVR-34B generates visual reasoning traces directly, built on ByteDance's Emu3.5 with two-stage SFT (…

Introduces Direct On-Policy Distillation, a weak-to-strong post-training method that transfers the change induced by reinforcement learning rather tha…

Multi-scale autoregressive model for protein backbone generation: coarse-to-fine next-scale prediction with a flow decoder, enabling zero-shot motif s…

Benchmark of 134 day-long real-world agent tasks (51 publicly released) measuring how much agents improve by learning from their environment over 12&n…

ByteDance's flagship refresh, announced at the 2026 Volcano Engine Force Conference in Pro and Turbo tiers. ByteDance positions Pro as crossing a "pro…

"Continuous Latent Diffusion Language Model." A hierarchical latent diffusion LM that replaces token-level autoregression with continuous-latent prior…

Post-training method for aligning flow matching generative models with human preferences at any generation step count. Compresses long ODE trajectorie…

"Same Pretraining Loss, Better Downstream Generalization via Common Minima." Proposes that the geometric closeness of task-specific minima during pret…

"Evidence From Competitive Programming." Empirically demonstrates a log-linear relationship between accuracy and the number of reasoning tokens genera…

Felis is an open-source, automated, and scalable toolkit designed for high-throughput Absolute Binding Free Energy (ABFE) calculations. Paired with By…

Novel attention mechanism where each attention head can attend to both sequence KV pairs at the current layer and depth KV pairs from preceding layers…

ByteDance's frontier general-purpose family, announced February 14, 2026. Architecture is "systematically optimized" for long-chain reasoning and agen…

An open-source, trainable reproduction of AlphaFold 3 (AF3) for biomolecular structure prediction. Protenix is designed to predict the 3D structures o…

Native multimodal audio-video joint generation model accepting four input modalities (text, image, audio, video) and producing 4–15 second clips…

"Post-LayerNorm Is Back: Stable, ExpressivE, and Deep." Revisits Post-LN Transformers — previously abandoned in favor of Pre-LN due to training…

1.5 trillion-token "distilled" subset of common crawl data optimized for pre-training.

A next-generation biomolecular folding model that scales up model capacity and performance beyond AlphaFold3. SeedFold introduces a width-scaling stra…

Generalized agentic model with multimodal input (text + image). Industry-leading GUI agent, search, and code generation capabilities across desktop, w…

ByteDance's frontier coding-focused reasoning model, released November 2025. Reasoning model with extended chain-of-thought; multimodal text+image inp…

Unified spacetime autoregressive modeling framework for visual generation. Extends autoregressive approaches to jointly model spatial and temporal dim…

Named after the recursive Ouroboros, Ouro is an open-source family of Looped Language Models (LoopLM) that embed reasoning during pretraining rather t…

Proposes Parallel Loop Transformer (PLT), which eliminates the sequential bottleneck of looped Transformers by computing different loops for different…

Foundation model for generating simulation-ready 3D assets from single images.

ByteDance Seed (with CUHK-Shenzhen and the Shenzhen Research Institute of Big Data; first author Ziniu Li, senior author Zhi-Quan Luo) attacks the deg…

Robotic wrist design and learning framework for flexible, anthropomorphic motion.

Unified framework for text-to-image, image editing, and multi-image composition. Ranked first on both T2I and image-editing leaderboards.

Open-source framework for scaling any-modality foundation model training.

Industry-level video foundation model for unified text-to-video (T2V) and image-to-video (I2V) generation.

Open-source 36B-parameter LLM with 512K context window. Achieves strong performance across reasoning, agent, and general tasks despite being trained o…

A generative protein binder design suite featuring diffusion-based backbone generation (PXDesign-d) and hallucination-based methods (PXDesign-h). Vali…

Research on scaling any-modality training recipes. Accepted at AAAI 2026.

Experimental diffusion-based language model capable of inference speeds exceeding 2,100 tokens/s.

Specialized model for automated theorem proving and formal mathematical reasoning (Lean 4).

Open-source multilingual translation series (7B parameters) rivaling GPT-4 in translation tasks. Includes Seed-X-Instruct, Seed-X-PPO, and Seed-X-RM.

General-purpose multimodal model series introducing Adaptive Chain-of-Thought (AdaCoT) — a reward-shaped RL recipe that teaches the model to dec…

ByteDance's video generation model. Ranked first on Artificial Analysis leaderboards for both text-to-video and image-to-video. Generates 5-second 108…

14B MoE unified multimodal model integrating vision-understanding and image-generation into a single framework.

Research on model merging techniques during pre-training of large language models, exploring how independently trained model branches can be merged to…

Vision-language foundation model with a 532M vision encoder and 20B-active MoE LLM. State-of-the-art on 38 of 60 public VLM benchmarks. Excels at GUI…

Open-source 8B code-LLM family (previously branded "Doubao-Coder"). Three released variants: Base (32K), Instruct (32K), and Reasoning (64K, RL-traine…

"SuperAgent" harness that orchestrates sub-agents and tools for long-horizon research and coding tasks.

RL framework for teaching LLMs strategic tool use by dynamically interleaving real-time code execution within natural language reasoning. Uses outcome…

Reasoning model advanced through reinforcement learning.

Reasoning-focused model that surpassed OpenAI's o1-preview on the AIME benchmark.

A systematic study of "zero RL training" — applying rule-based-reward reinforcement learning directly to base models (the DeepSeek-R1 paradigm) — acro…

Decoupled Clip and Dynamic sAmpling Policy Optimization — an open-source LLM reinforcement-learning system at scale, from the SIA-Lab (the joint…

Native bilingual Chinese-English image generation model with integrated LLM text encoder and character-level text rendering.

Unified tokenizer for visual generation and understanding. Uses multi-codebook quantization to scale vocabulary size and bottleneck dimension, achievi…

The most comprehensive graduate-level knowledge reasoning benchmark, spanning 285 disciplines (13 major, 72 fields, 285 subfields) with 26,529 questio…

Argues that a Transformer's input and output vocabularies should be decoupled and scaled independently. The Over-Tokenized Transformer (OT) pairs Over…

Sparse MoE model (200B total, 20B active, 256 experts with 8 active per token) with a 256k context window. Part of ByteDance's flagship Doubao series.

Open-source multimodal AI agent stack designed to control computers and browsers via vision.

Scaling bitwise autoregressive modeling for high-resolution image and spacetime (video) synthesis.

Unified multimodal generation framework that combines visual understanding and image generation within a single LLM without external pretrained visual…

Unified optimization framework that brings variance reduction back into contention for large-model training, where AdamW has dominated for a decade. M…

Applying the Llama architecture to image generation.

State-of-the-art speech generation model family known for high-fidelity voice cloning and expressive prosody.

An Amber-compatible, data-driven force field for drug-like molecules. ByteFF is trained on millions of synthetic fragment geometries using an edge-aug…

NeurIPS 2024 Best Paper. Paradigm shift proving GPT-style autoregressive models could surpass diffusion (SDXL/Stable Diffusion) in image generation qu…

System for scaling LLM training to over 10,000 GPUs.

Lightning-fast distillation of SDXL allowing high-quality image generation in 1-4 steps.

Highly popular foundation model for monocular depth estimation.

Viral diffusion-based models for human image animation and high-dynamic video generation.

High-performance Transformer implementation optimized for variable-length sequences.

Legendary multi-object tracking algorithm that remains an industry standard.