ByteDance

SeedRealtime

ByteDance

SeedRealtime is ByteDance’s native full-duplex audio-visual language model for continuous conversation. It unifies text, audio, video, conversational timing, and response generation in one architecture. Streaming perception allows it to use the user’s environment to resolve ambiguous speech and maintain context in noisy multi-speaker scenes. The model was publicly described as fully rolled out in ByteDance products. Future tool-driven actions remain roadmap work; the launch does not establish downloadable weights or a general developer API.

Doubao Seed 2.1 Pro

ByteDance

Doubao Seed 2.1 Pro is ByteDance’s flagship productivity model in the Seed 2.1 family. It combines language reasoning with image and video understanding for document-heavy tasks, visual analysis, and multi-step agents. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

Doubao Seed 2.1 Turbo

ByteDance

Doubao Seed 2.1 Turbo is ByteDance’s efficient productivity tier in the Seed 2.1 family. It combines language reasoning with image and video understanding for document-heavy tasks, visual analysis, and multi-step agents. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

Doubao Seed 2.0 Code

ByteDance

Doubao Seed 2.0 Code is ByteDance’s dedicated software engineering model in the Seed 2.0 family. The release strengthens repository changes, debugging, and code delivery through an external agent environment. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

Doubao Seed 2.0 Lite

ByteDance

Doubao Seed 2.0 Lite is ByteDance’s efficient productivity tier in the Seed 2.0 family. It combines language reasoning with image and video understanding for document-heavy tasks, visual analysis, and multi-step agents. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

Doubao Seed 2.0 Mini

ByteDance

Doubao Seed 2.0 Mini is ByteDance’s efficient productivity tier in the Seed 2.0 family. It combines language reasoning with image and video understanding for document-heavy tasks, visual analysis, and multi-step agents. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

Doubao Seed 2.0 Pro

ByteDance

Doubao Seed 2.0 Pro is ByteDance’s flagship productivity model in the Seed 2.0 family. It combines language reasoning with image and video understanding for document-heavy tasks, visual analysis, and multi-step agents. Public inference is offered through Volcengine and ByteDance applications, with no downloadable weights. The tier is a separate product identity; API snapshots and later service migrations are variants. Parameter counts and unsupported endpoint limits are left unspecified.

Stable DiffCoder

ByteDance

Stable DiffCoder is ByteDance’s 8B diffusion coding model, built on the Seed Coder architecture and training pipeline. Block-diffusion continual pretraining uses a tailored warmup and clipped noise schedule, followed by instruction tuning. It generates code through iterative parallel refinement and supports code completion, editing, and reasoning. MIT-licensed base and instruction weights and the training report enable controlled comparisons against autoregressive coding models. Its documented 8K context and specialized inference routine should be respected instead of assuming ordinary sequential decoding behavior.

Doubao Seed 1.8

ByteDance

Seed1.8 is ByteDance’s generalized agentic language and vision model, served through Volcano Engine as Doubao 1.8 and through BytePlus. It integrates search, code generation, and graphical-interface reasoning, with three thinking modes for tasks of varying complexity. Text and image inputs support language answers and sustained workflows, including educational questions, information processing, and software engineering. Its official model portfolio dates the release to December 17, preceding the December 18 explanatory announcement. Public weights, parameter count, and full backbone details are not disclosed in the reviewed sources.

UI-TARS 2

ByteDance

UI-TARS 2 is ByteDance’s native GUI-centered research model, first published in its September 2 technical report. It starts from Seed1.6 checkpoints and combines a 230B-parameter mixture-of-experts language backbone, activating 23B parameters, with a 532M vision encoder. Multi-turn reinforcement learning trains language reasoning and actions across GUI, game, terminal, file-system, and tool environments. Parameter interpolation combines specialized agents into a broader model. The published research and demonstration qualify under the historical research-model rule, but reviewed sources do not establish publicly downloadable UI-TARS 2 weights or a general public model API.

Seed OSS 36B

ByteDance

Seed OSS is ByteDance’s open language model family with a 36B dense backbone and native 512K context. It supports controllable thinking budgets, reasoning, and tool-using agents, targeting international developer use cases. The release groups instruction and base checkpoints, including base variants with and without synthetic training data. Apache-licensed weights allow private inference and fine-tuning. Benchmark results depend on reasoning budgets and deployment settings, while the full long-context capacity can require considerably more memory than short-prompt serving.

Doubao Seed 1.6 Flash

ByteDance

Seed1.6 Flash is ByteDance’s low-latency multimodal reasoning tier, served through Volcano Engine and BytePlus. It supports text and visual input, optional deep thinking, and a documented 256K context with up to 16K output tokens. The model targets responsive assistants and high-volume visual or language tasks while retaining reasoning capability. Its independently positioned Flash tier is represented separately from the larger Seed1.6 model; its parameter count and detailed architecture are not publicly established in the reviewed sources. June 2025 release-month precision follows the official API catalog, without treating the 250615 snapshot suffix as a launch date.

Doubao Seed 1.6 Thinking

ByteDance

Seed1.6 Thinking is ByteDance’s separately named extended-reasoning tier for mathematical, coding, and visual problems. It is post-trained on the Seed1.6 mixture-of-experts foundation with repeated rejection fine-tuning and reinforcement learning, integrating image understanding into long reasoning traces. It emphasizes sustained deliberation rather than selecting thinking automatically for each prompt. Volcano Engine offers the model under Doubao API identifiers, with a 256K context. The official June 2025 technical introduction documents a public try link; the release month is retained because the first public day is not independently verified.

Doubao Seed 1.6

ByteDance

Seed1.6 is ByteDance’s multimodal foundation model with adaptive reasoning, offered under the Doubao name through Volcano Engine and internationally through BytePlus. Its 230B-parameter mixture-of-experts backbone activates 23B parameters and supports a 256K context. AdaCoT post-training lets callers choose automatic, full, or disabled thinking to balance latency and reasoning. Text and images can inform language answers, coding, document analysis, and interface understanding. Public weights are not offered. June 2025 availability is established by the official technical introduction and API catalog; the exact first public launch day is not asserted here.

BAGEL

ByteDance

BAGEL is ByteDance Seed’s unified language-capable multimodal model for visual understanding, text generation, image creation, and editing. It uses a mixture of Transformer experts with 14B total and 7B active parameters, combining semantic and pixel-level image encoders with interleaved multimodal training. Image questions and language reasoning make it eligible alongside its generative visual capabilities. Apache-licensed weights and inference code were released on May 20, before the later May 28 Seed blog. The model is fine-tuned from Qwen2.5-7B-Instruct with additional vision components, and can be run locally or evaluated through its demo.

Seed-Coder

ByteDance

Seed-Coder is ByteDance Seed’s 8B coding language-model family, released on May 8 under MIT. Its base, instruction, and reasoning checkpoints share a model-centric data pipeline in which language models curate code corpora rather than relying solely on handcrafted filters. Base and instruction models support 32K context, while the reinforcement-trained reasoning variant supports 64K; this page uses the common 32K limit and records the larger variant explicitly. The base supports fill-in-the-middle completion, while instruction and reasoning variants address conversational code generation and difficult programming tasks. All routine checkpoints are grouped on one family page.

UI-TARS 1.5

ByteDance

UI-TARS 1.5 is ByteDance’s reasoning-focused upgrade to its native GUI vision-language model. The released 7B checkpoint uses reinforcement learning to produce deliberative language traces before interface actions, improving adaptability on desktop, browser, and game tasks. The official repository records April 16 publication and open-weight availability. Apache-licensed weights are available for local deployment, while the UI-TARS desktop application supplies execution and interaction machinery separately. This page represents the named model update rather than the application. Screenshots and text are inputs; language reasoning and grounded action instructions are outputs.

UI-TARS

ByteDance

UI-TARS is ByteDance’s native vision-language model family for computer interaction. It reads screenshots, reasons in language, and generates grounded actions, integrating perception, planning, and memory in the model rather than relying on a fixed workflow. The initial paper was published on January 21, with Apache-licensed 2B, 7B, and 72B checkpoints in supervised and preference-trained variants. These routine sizes are grouped on one family page. Downloadable weights support local experimentation with browser, desktop, and mobile agents; the desktop application and execution harness are separate components that carry out the model’s proposed actions.