Doubao Seed is the family of large language models built by ByteDance's Seed team, served in China through the Doubao app and Volcano Engine and internationally through the BytePlus ModelArk API and the Dola app; its newest generation, Seed 2.1, was announced on 23 June 2026. This page lists the Seed language models that ByteDance documents on its official pages, with API model IDs, context windows, published BytePlus prices in US dollars, release dates, retirement dates and the one open-weight Seed language model. It is written for developers, buyers and architects who need to know which Seed model is which. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from ByteDance Seed or BytePlus documentation, cited in the sources list. ByteDance's video, image, audio, robotics and science models are out of scope apart from a one-line mention.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses ByteDance Seed's public model pages and BytePlus ModelArk documentation as of November 2026. Model details change, so confirm against ByteDance's current documentation.
Key takeaways
- Seed 2.1 is the current generation. ByteDance Seed's blog dated 23 June 2026 announces Seed 2.1 with access for Doubao and Volcano Engine users, and the Seed 2.1 page names two sizes, Pro and Turbo; BytePlus ModelArk lists dola-seed-2-1-turbo-260628.
- Seed 2.0 remains on the API. Launched on 14 February 2026 as Pro, Lite and Mini agent models plus a Code model, its dated IDs seed-2-0-pro-260328, seed-2-0-lite-260428, seed-2-0-mini-260428 and seed-2-0-code-preview-260328 are on the ModelArk model list.
- Every listed Seed model has a 256K context. The ModelArk model list gives 256K context for all five Seed LLM IDs, with 256K max output for Seed 2.1 Turbo and 128K for the Seed 2.0 models; default output is 4K unless set.
- BytePlus prices run from $0.10 to $1.00 per million input tokens. Seed 2.0 Mini starts at $0.10 input and $0.40 output; Seed 2.1 Turbo is $0.50 input and $2.50 output; prices step up above 128K prompt tokens for the Seed 2.0 models, and Flex rates are half of Standard input and output.
- Retirements come in dated batches. ModelArk deactivated the Seed 1.6, Seed 1.8 and early Seed 2.0 snapshots on 11 November 2026 and states at least 30 calendar days' notice before deactivation.
What is ByteDance Seed?
ByteDance Seed is the research team that builds ByteDance's foundation models. Its model page groups work into general agents and coding (the Seed language models), video (Seedance), image (Seedream), real-time interaction (SeedRealtime), audio (Seed Audio), robotics (GR-3, GR-RL) and AI for science (Protenix). This page covers the language models only.
The same models appear under several names. In China the consumer chatbot is the Doubao app and the enterprise API is Volcano Engine; the Seed 2.1 blog refers to "Doubao Seed 2.1". Outside China the API is BytePlus ModelArk, a ByteDance cloud brand, and the consumer chat app linked from the Seed pages is Dola. Seed 2.1 model IDs on BytePlus carry a "dola-" prefix. The Seed language models served on these platforms are proprietary; ByteDance publishes one open-weight language model series, Seed-OSS-36B, under Apache 2.0. Facts are as of 15 November 2026.
Current Seed models at a glance
As of 2026-11-15. Source: ByteDance Seed model and blog pages, BytePlus ModelArk model list. Models are listed by release date, newest first. Category is the Seed generation.
| Model | API ID or version | Category | Context | Licence | Status |
|---|---|---|---|---|---|
| Seed 2.1 Turbo | dola-seed-2-1-turbo-260628 | Seed 2.1 | 256K | Proprietary | Active on BytePlus ModelArk |
| Seed 2.1 Pro | Not stated by ByteDance | Seed 2.1 | Not stated by ByteDance | Proprietary | Announced; Doubao and Volcano Engine access per blog |
| Seed 2.0 Lite | seed-2-0-lite-260428 | Seed 2.0 | 256K | Proprietary | Active |
| Seed 2.0 Mini | seed-2-0-mini-260428 | Seed 2.0 | 256K | Proprietary | Active |
| Seed 2.0 Pro | seed-2-0-pro-260328 | Seed 2.0 | 256K | Proprietary | Active |
| Seed 2.0 Code (preview) | seed-2-0-code-preview-260328 | Seed 2.0 | 256K | Proprietary | Active, preview ID |
| Seed-OSS-36B | Base, Base-woSyn, Instruct | Open weights | 512K | Apache 2.0 | Published on Hugging Face |
The ModelArk model list marks every Seed LLM ID as supporting deep reasoning, visual understanding and function calling. Audio understanding is listed for Seed 2.0 Lite and Mini only. The Seed 2.0 page describes Lite as the first omni-modal model in the Seed foundation series, handling video, image, audio and text, with audio added in a late-April update. See Open-Weight vs Closed Models Explained.
Key facts for the main Seed models
As of 2026-11-15. Identical fields for every model; "Not stated by ByteDance" means not stated by ByteDance on the pages checked (2026-11-15).
| Model | Released | Max output | Knowledge cutoff | Input / output | Open weights | Parameters | Positioning (provider wording) |
|---|---|---|---|---|---|---|---|
| Seed 2.1 Turbo | 2026-06-23 (family) | 256K (includes reasoning) | Not stated by ByteDance | Text, image / text | No | Not stated by ByteDance | AI productivity model, smaller of two sizes |
| Seed 2.1 Pro | 2026-06-23 (family) | Not stated by ByteDance | Not stated by ByteDance | Not stated by ByteDance | No | Not stated by ByteDance | AI productivity model, larger of two sizes |
| Seed 2.0 Pro | 2026-02-14 (family) | 128K (includes reasoning) | Not stated by ByteDance | Text, image / text | No | Not stated by ByteDance | Long-chain reasoning and robustness in complex workflows |
| Seed 2.0 Lite | 2026-02-14 (family) | 128K (includes reasoning) | Not stated by ByteDance | Text, image, video, audio / text | No | Not stated by ByteDance | Balances output quality and response speed |
| Seed 2.0 Mini | 2026-02-14 (family) | 128K (includes reasoning) | Not stated by ByteDance | Text, image, audio / text | No | Not stated by ByteDance | Inference throughput and high-concurrency batch generation |
| Seed 2.0 Code | 2026-02-14 (family) | 128K (includes reasoning) | Not stated by ByteDance | Text, image / text | No | Not stated by ByteDance | Dedicated code model |
| Seed-OSS-36B-Instruct | 2025-08-20 | Not stated by ByteDance | Not stated by ByteDance | Text / text | Yes | 36B | Reasoning, agentic tasks and long context |
Dated suffixes in the API IDs (for example 260428) are version identifiers. The Seed pages do not state that they are release dates, so this page does not treat them as such.
Seed model families and naming
- Generation number: Seed 1.6, Seed 1.8, Seed 2.0 and Seed 2.1 are successive generations. Seed 1.6 and 1.8 IDs were deactivated on BytePlus on 11 November 2026.
- Size tiers: Seed 2.0 uses Pro, Lite and Mini for general agent models plus a separate Code model. Seed 2.1 uses Pro and Turbo.
- Snapshot suffix: BytePlus IDs end in a six-digit snapshot such as -260328 or -260628. A new snapshot of the same model gets a new ID, and older snapshots are retired (seed-2-0-lite-260228 was replaced by seed-2-0-lite-260428).
- Prefixes: Seed 2.1 IDs use "dola-seed"; Seed 2.0 IDs use "seed". Older Doubao 1.5 IDs used "doubao", and an earlier "skylark" line was retired in 2025 and 2026.
- Brand names: Doubao (China app and model brand), Dola (international app), Volcano Engine (China cloud), BytePlus ModelArk (international cloud). The model underneath is the same Seed generation.
- Open weights: Seed-OSS-36B is a separate open-weight series, not a size tier of Seed 2.x.
When was each Seed model released?
As of 2026-11-15. Dates are from ByteDance Seed blog and Hugging Face pages and BytePlus deprecation notices.
| Date | Event | API ID |
|---|---|---|
| 2026-11-11 | Fourth deprecation batch deactivated (Seed 1.6, Seed 1.8, early Seed 2.0 snapshots) | seed-1-6-250915, seed-1-8-251228 and others |
| 2026-09-15 | Fourth batch deprecated; no new endpoints from this date | As above |
| 2026-06-23 | Seed 2.1 Pro and Turbo announced (blog date) | dola-seed-2-1-turbo-260628 |
| 2026-05-13 | Doubao 1.5 Pro 32K deactivated on BytePlus | doubao-1-5-pro-32k-250115 |
| 2026-02-14 | Seed 2.0 Pro, Lite, Mini and Code launched (blog date) | Later snapshots: seed-2-0-pro-260328 and others |
| 2025-08-20 | Seed-OSS-36B Base, Base-woSyn and Instruct released on Hugging Face | Open weights |
The Seed 2.1 blog describes general agent work across tools, computer use (reporting that reinforcement learning cut average steps per task by 16%), full-cycle coding with coordinated multi-file changes, and internal use in a "Seed for Seed" multi-agent research initiative. The Seed 2.1 page says the family is "officially released" but gives no separate date for Turbo and Pro. For dated release history across providers see the LLM Release Tracker.
How much do Seed models cost?
As of 2026-11-15. Source: BytePlus ModelArk model pricing page (no last-updated date shown). USD per million tokens, Standard mode, prompt length up to 128K tokens unless noted. Cache storage is $0.0083 per million tokens per hour for every row.
| Model | Input | Cached input | Output | Batch | Cache write | Notes |
|---|---|---|---|---|---|---|
| dola-seed-2-1-turbo | $0.50 | $0.10 | $2.50 | Not stated by ByteDance | Storage $0.0083 per M per hour | Single tier 0 to 256K; Flex $0.25 / $1.25 |
| seed-2-0-pro-260328 | $0.50 | $0.10 | $3.00 | Not stated by ByteDance | Storage $0.0083 per M per hour | Above 128K: $1.00 / $0.20 / $6.00; Flex half input and output |
| seed-2-0-code-preview-260328 | $0.50 | $0.10 | $3.00 | Not stated by ByteDance | Storage $0.0083 per M per hour | Above 128K: $1.00 / $0.20 / $6.00 |
| seed-2-0-lite-260428 | $0.25 | $0.05 | $2.00 | $0.125 / $1.00 | Storage $0.0083 per M per hour | Above 128K: $0.50 / $0.10 / $4.00; audio input $3.75 |
| seed-2-0-mini-260428 | $0.10 | $0.02 | $0.40 | $0.05 / $0.20 | Storage $0.0083 per M per hour | Above 128K: $0.20 / $0.04 / $0.80; audio input $1.50 |
The pricing page sets the tier by each request's prompt length: a 200K-token prompt is billed at the (128, 256]K rate. Flex mode charges half the Standard input and output rates with the same cached-input price and no cache-storage fee. The page lists no free quota for these models. Seed 2.1 Pro has no BytePlus price row. Volcano Engine prices in yuan for the China market were not retrievable from an official page and are not shown. Compare providers in the LLM Pricing Tracker.
Where are Seed models available?
As of 2026-11-15.
| Channel | Details | Source |
|---|---|---|
| BytePlus ModelArk API | All five Seed LLM IDs in ap-southeast-1; Seed 2.0 models also in eu-west-1 | ModelArk model list |
| BytePlus Playground | Try-it links for Seed 2.1 Turbo and Seed 2.0 Pro | Seed 2.1 and Seed 2.0 pages |
| Volcano Engine | Seed 2.0 full-series API; Seed 2.1 for Volcano Engine users | Seed 2.0 and 2.1 blogs |
| Doubao app | Seed 2.0 Pro and Code; Seed 2.1 for Doubao users | Seed 2.0 and 2.1 blogs |
| Dola app | Linked as "Try Dola" from the Seed 2.0 and 2.1 pages | Seed model pages |
| TRAE | Seed 2.0 Pro and Code | Seed 2.0 blog |
| Hugging Face | Seed-OSS-36B weights | Seed-OSS model card |
The ModelArk model list states per-model rate limits: 0.5K requests per minute and 1,000K tokens per minute for dola-seed-2-1-turbo-260628, and 30K requests per minute and 1,500K tokens per minute for the Seed 2.0 IDs.
Restricted-access Seed models
| Model | Access stated by provider | Data retention condition |
|---|---|---|
| Seed 2.1 Pro | Doubao and Volcano Engine users (blog); no BytePlus model ID or price row on the pages checked | Not stated by ByteDance |
| All other listed models | Generally available on the channels above; no restricted-access tier stated | Content-filter-triggered data retained 180 days (BytePlus) |
What is the context window of Seed models?
All five Seed LLM IDs on the ModelArk model list have a 256K-token context. Seed 2.1 Turbo accepts up to 256K input and 256K output, with up to 256K of chain of thought. The four Seed 2.0 IDs accept up to 224K input and 128K output, with up to 128K of chain of thought. The model list states that max output includes chain of thought and that the default output is 4K tokens unless set, so long answers need an explicit max-tokens value. Seed-OSS-36B is trained natively with up to 512K context, according to its model card. Seed 2.1 Pro's context window is not stated on the pages checked.
Is any Seed model open source?
The Seed 2.x models are proprietary and served only through ByteDance channels. ByteDance publishes Seed-OSS-36B as open weights under Apache 2.0. The model card lists three checkpoints released on 20 August 2025: Seed-OSS-36B-Base (trained with synthetic instruction data), Seed-OSS-36B-Base-woSyn (without synthetic instruction data) and Seed-OSS-36B-Instruct. The card states training on 12T tokens and a user-set thinking budget, with multiples of 512 recommended and 0 for a direct answer. Running open weights moves security, patching and access control to the operator; see the Open-Weight Model Security Guide and the Open-Weight Model Licence Tracker.
Model cards and system cards
| Model | Model card or page | System card | Technical report |
|---|---|---|---|
| Seed 2.1 | Seed 2.1 page with model card PDF link | Not stated by ByteDance | Seed 2.1 tech blog |
| Seed 2.0 | Seed 2.0 page with model card PDF link | Not stated by ByteDance | Seed 2.0 tech blog |
| Seed-OSS-36B | Hugging Face model card | Not stated by ByteDance | Not stated by ByteDance |
The Seed 2.1 model card PDF could not be retrieved on the check date; facts on this page come from the Seed 2.1 web page and blog. Neither the Seed 2.0 nor the Seed 2.1 blog has a safety section. See Model Cards and System Cards Explained.
Benchmarks
Provider-reported scores, as published on the Seed 2.1 and Seed-OSS pages (checked 2026-11-15). ByteDance chose the benchmarks, settings and comparators; scores are not independently verified. Seed 2.0 scores are in the Seed 2.0 model card images and are not reproduced here.
| Benchmark | Seed 2.1 Pro (provider-reported) | Seed 2.1 Turbo (provider-reported) | Independent leaderboard results |
|---|---|---|---|
| SuperGPQA | 70.8 | 67.4 | Not checked |
| BeyondAIME | 87.0 | 88.0 | Not checked |
| Terminal Bench 2.1 | 71.0 | 67.6 | Not checked |
| SWE-Atlas | 35.2 | 30.6 | Not checked |
| NL2Repo-Bench | 47.0 | 43.7 | Not checked |
| MMMU-Pro (82.7 Pro and 82.2 Turbo with tools) | 81.6 | 80.1 | Not checked |
| VideoMME | 89.2 | 89.0 | Not checked |
| MMLongBench-128K | 78.3 | 76.9 | Not checked |
The Seed 2.1 blog reports that a Seed 2.1 Preview placed 8th on the Code Arena Frontend leaderboard with a score of 1539, and the Seed 2.0 blog reports 6th on LMSYS Text Arena overall as of 16 February 2026; these are ByteDance's statements of third-party rankings. The Seed-OSS-36B-Instruct card reports AIME25 84.7, LiveCodeBench v6 67.4, SWE-Bench Verified (OpenHands) 56 and MMLU-Pro 82.7. Cross-provider context is in Frontier AI Models Compared.
Deprecations and retirements
As of 2026-11-15. Source: BytePlus ModelArk model deprecation notice. After the deprecation date no new inference endpoints or fine-tuning tasks can be created; on the deactivation date the model stops serving.
| Model | API ID | Deprecated or notice | Retirement or shutdown | Stated replacement |
|---|---|---|---|---|
| Seed 1.8 | seed-1-8-251228 | 2026-09-15 | 2026-11-11 | seed-2-0-lite-260428 |
| Seed 1.6 | seed-1-6-250615, seed-1-6-250915 | 2026-09-15 | 2026-11-11 | seed-2-0-lite-260428 |
| Seed 1.6 Flash | seed-1-6-flash-250615, seed-1-6-flash-250715 | 2026-09-15 | 2026-11-11 | seed-2-0-mini-260428 |
| Seed 1.6 Code preview | seed-1-6-code-preview | 2026-09-15 | 2026-11-11 | dola-seed-2-1-turbo-260628 |
| Seed 2.0 Lite (early snapshot) | seed-2-0-lite-260228 | 2026-09-15 | 2026-11-11 | seed-2-0-lite-260428 |
| Seed 2.0 Mini (early snapshot) | seed-2-0-mini-260215 | 2026-09-15 | 2026-11-11 | seed-2-0-mini-260428 |
| Doubao 1.5 Pro 32K | doubao-1-5-pro-32k-250115 | 2026-03-23 | 2026-05-13 | seed-2-0-mini-260215 |
| Skylark Pro | skylark-pro-250415 | 2025-11-24 | 2026-01-22 | seed-1-6-250915 |
The notice states that notifications are issued at least 30 calendar days before deactivation and that, in general, each model version is deprecated and deactivated a few months after launch. Teams should plan for snapshot changes several times a year.
Data handling and security
- Training use: the BytePlus ModelArk data processing page states that BytePlus will not use Customer Data for its own model training without the customer's prior authorisation.
- Retention: input and output that trigger the content filter are retained for 180 days in Malaysia for content safety purposes. A zero-retention option is not stated by ByteDance on the pages checked.
- Locations: model deployment and processing may use data centres in Malaysia, Indonesia and the EU/EEA; the model list names ap-southeast-1 and eu-west-1.
- Encryption: traffic uses HTTPS, and an optional application-layer end-to-end encryption feature encrypts data with the inference instance's certificate public key.
- Jurisdiction: Doubao and Volcano Engine are Chinese services; see the China Generative AI Measures and AI Labelling Rules Guide and China Cross-Border Data Transfers.
- Credentials: ModelArk API keys are non-human identities with spend and data exposure; see the LLMjacking Guide. Provider policies side by side are in LLM Data Retention and Training-Use Policies by Provider.
How to choose a Seed model: factual criteria
| Requirement | Documented fact to check |
|---|---|
| Newest generation on BytePlus | dola-seed-2-1-turbo-260628, the only Seed 2.1 ID on the model list |
| Output above 128K tokens | Seed 2.1 Turbo lists 256K max output; Seed 2.0 IDs list 128K |
| Audio input | Seed 2.0 Lite and Mini list audio understanding |
| Lowest published token price | seed-2-0-mini-260428 at $0.10 input, $0.05 in Flex or Batch |
| EU region processing | Seed 2.0 models are listed for eu-west-1 |
| High request rates | Seed 2.0 IDs list 30K RPM; Seed 2.1 Turbo lists 0.5K RPM |
| Weights you can host | Seed-OSS-36B under Apache 2.0 |
Evaluate shortlisted models on your own data; provider benchmark tables use ByteDance's chosen settings and comparators.
Practitioner checklist
- Pin the full dated model ID (for example seed-2-0-lite-260428) and record the date adopted.
- Subscribe to ModelArk deprecation notices: snapshots retire within months, with at least 30 days' notice.
- Set max output explicitly; the default is 4K tokens and max output includes reasoning tokens.
- Budget for the higher tier when prompts exceed 128K tokens on Seed 2.0 models.
- Use Flex or Batch for workloads that tolerate delay, where the pricing page halves input and output rates.
- Choose the region (ap-southeast-1 or eu-west-1) to match your data residency needs, and confirm content-filter retention is acceptable.
- Store ModelArk API keys in a secrets manager, scope them per application and set spend alerts.
- Run your own evaluations; provider-reported scores are not independent.
Standards and references
- ByteDance Seed: Models (accessed 2026-11-15)
- ByteDance Seed: Seed 2.1 (accessed 2026-11-15)
- ByteDance Seed: Seed 2.1 officially released (accessed 2026-11-15)
- ByteDance Seed: Seed 2.0 (accessed 2026-11-15)
- ByteDance Seed: Seed 2.0 official launch (accessed 2026-11-15)
- BytePlus: ModelArk model list (accessed 2026-11-15)
- BytePlus: ModelArk model pricing (accessed 2026-11-15)
- BytePlus: ModelArk model deprecation notice (accessed 2026-11-15)
- BytePlus: ModelArk data processing (accessed 2026-11-15)
- Hugging Face: ByteDance-Seed/Seed-OSS-36B-Instruct model card (accessed 2026-11-15)
Related NHI Mgmt Group resources: Tencent Hunyuan Models, Alibaba Qwen Models, DeepSeek Models, Open-Weight Frontier Models Compared.