Latest5 Aug 2026Audio/voice model
SeedRealtime
A full-duplex audio-visual model that watches, listens and talks over continuous streams in real time, instead of chaining separate speech and vision modules. ByteDance says it decides its own conversational timing and can speak up unprompted. It launched inside the Doubao app.
- Native audio-visual full-duplex interaction
- Uses what it sees to disambiguate speech
- Tells background noise from speech meant for it
- Speaks proactively based on its surroundings
AccessDoubao app (free, per press); Seed blog says it is fully rolled out; no API disclosed
31 Jul 2026Video model
Seedance 2.5
ByteDance's newest joint audio-video generation model. It makes clips of up to 30 seconds in one run, which can be extended twice, and takes many reference images, videos and audio clips at once. Green-screen and white-model (blocking) controls let users stage scenes and camera moves.
- Up to 30 s per generation, extendable twice
- Joint audio and video generation with audio editing
- Up to 30 images, 10 videos, 10 audio refs (press)
- Green-screen and white-model staging control
AccessJimeng/Dreamina, Doubao; Volcano Engine Ark and BytePlus (API announced as coming); Dola
8 Jul 2026Image model
Seedream 5.0 Pro
ByteDance's top image generation and editing model, aimed at text-heavy professional graphics such as infographics and annotated layouts. It accepts edits marked with points, circles or sketches, separates layers, and renders text natively in about a dozen languages. It is offered in Doubao, Jimeng and through ByteDance's cloud APIs.
- Dense infographics and information layouts
- Edits guided by points, circles and sketches
- Native text rendering in ~12 languages
- Photographic lighting, material and skin detail
AccessVolcano Engine Ark, Doubao, Jimeng/Dreamina; BytePlus API; Dola
24 Jun 2026Frontier LLM
Seed2.1 Pro
ByteDance's current flagship language model, released together with a cheaper Seed2.1 Turbo tier. It is aimed at coding, long-running agent tasks and multimodal understanding, including images and hour-long video. ByteDance compares its coding and agent results with leading Western models.
- Text, image and video input, incl. hour-long video
- Pro and Turbo tiers (Turbo half price, per press)
- Lab benchmarks incl. Terminal Bench 2.1, SWE-Atlas, MMMU-Pro
- Long-context evaluation to 128K (MMLongBench-128K)
AccessDola/Doubao chat; Volcano Engine Ark and BytePlus ModelArk API; BytePlus playground
14 Feb 2026Frontier LLM
Seed2.0 Pro
ByteDance's 2026 model generation, released in Pro, Lite and Mini tiers. Pro focuses on long-chain reasoning and complex workflows, Lite on balanced production use, and Mini on throughput. A later Lite upgrade made it ByteDance's first omni-modal understanding model, adding audio input.
- Pro / Lite / Mini tiers
- Video, image and text input (Lite adds audio)
- VideoMME 89.0 for Pro (lab-reported)
- Lite: SWE Multilingual 66.6% (lab-reported)
AccessDoubao app; Volcano Engine and BytePlus API; Dola chat
12 Feb 2026Video model
Seedance 2.0
A joint audio-video generation model that takes text, images, audio and video together as input. It gives director-style control over performance, lighting and camera movement, and produces synchronised sound and picture. It launched in ByteDance's Dreamina creative tools, with API access through BytePlus.
- Combined text, image, audio and video input
- Synchronised audio-video output
- Control of camera, lighting and performance
- API via BytePlus ModelArk
AccessJimeng/Dreamina (incl. inside CapCut); BytePlus API and playground; Dola
18 Dec 2025Frontier LLMReported
Seed1.8 (Doubao-Seed-1.8)
A multimodal Doubao model launched at Volcano Engine's FORCE conference in December 2025, focused on agent tool use and long-video analysis. Volcano Engine said video inference capacity doubled to 1,280 frames. The Seedance 1.5 Pro video model launched alongside it.
- Multimodal input incl. long video (1,280 frames)
- Agent and tool-invocation features
- Launched with AgentKit and HiAgent tooling
AccessDoubao app; Volcano Engine API
Nov 2025Coding modelReportedmonth only
Doubao-Seed-Code
Volcano Engine's model for agentic programming, integrated with ByteDance's Trae coding app. It handles 256,000 tokens per request and launched with a coding plan costing RMB 9.9 for the first month, which KrASIA put at about 63% below the market average. ByteDance reported 78.8% on SWE-Bench Verified.
- 256K-token context
- Agentic coding; integrated with Trae
- SWE-Bench Verified 78.8% (as reported)
- RMB 9.9 first-month coding plan
AccessVolcano Engine API and coding plan; works in ByteDance's Trae IDE