Doubao Seed 2.0 is a proprietary family of large language models from ByteDance's Seed team, launched on 14 February 2026 as three general agent models (Pro, Lite and Mini) plus a dedicated Code model. This profile collects the published facts about the family in one place: API identifiers, context window, output limits, prices on BytePlus ModelArk and in the Seed 2.0 model card, modalities, availability channels, provider-reported benchmarks, data handling and deprecation status. It is written for developers, architects, buyers and governance teams. Every fact comes from ByteDance Seed or BytePlus documentation, listed under Standards and references. Where ByteDance publishes nothing on a point, the page says so. Variants are listed in the order ByteDance uses (Pro, Lite, Mini, Code), which implies no ranking.
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 Seed 2.0 documentation and BytePlus ModelArk documentation as of November 2026. Model details change, so confirm against ByteDance's current documentation.
Key takeaways
- Launched 14 February 2026. The ByteDance Seed blog post "Seed 2.0 Official Launch" carries that date and introduces Pro, Lite and Mini as general agent models of different sizes, with a separate Code model.
- 256K context, 128K output on BytePlus. The ModelArk model list gives every Seed 2.0 ID a 256K-token context, 224K maximum input and 128K maximum output, with chain of thought counted inside the output limit.
- From $0.10 to $0.50 per million input tokens. BytePlus Standard prices for prompts up to 128K run from $0.10 input and $0.40 output (Mini) to $0.50 input and $3.00 output (Pro and Code), with higher rates above 128K.
- Superseded but active. Seed 2.1 (Pro and Turbo) is now ByteDance's newest generation, yet the current Seed 2.0 IDs remain on the ModelArk model list, and Lite and Mini 260428 are marked as LTS replacements.
- No system card. ByteDance publishes a Seed 2.0 model card PDF with capability benchmarks, but no separate system card or safety evaluation section was found on the pages checked.
What is Doubao Seed 2.0?
Seed 2.0 is the February 2026 generation of ByteDance's Seed language models. The same models appear under several names: Doubao is ByteDance's China consumer app and model brand, Volcano Engine is the China cloud that serves the API, and BytePlus ModelArk is the international API. The Seed 2.0 model card names the Pro model on Volcano Engine as "Doubao-Seed-2.0-pro", while BytePlus uses lower-case IDs such as seed-2-0-pro-260328.
The Seed 2.0 page positions the sizes as follows: Pro focuses on long-chain reasoning and robustness in complex workflows, Lite balances output quality and response speed, and Mini is optimised for inference throughput and deployment density for high-concurrency and batch generation. The launch blog describes the family as agent models with upgraded multimodal understanding across images, complex documents, tables, charts and video, including hour-long video. The models are proprietary and are served only through ByteDance channels. ByteDance does not state parameter counts. For the whole ByteDance line-up, including Seed 2.1 and the open-weight Seed-OSS-36B, see the ByteDance Doubao Seed Models.
Key facts at a glance
As of 2026-11-15. Sources: Seed 2.0 launch blog, Seed 2.0 page, Seed 2.0 model card, BytePlus ModelArk model list.
| Field | Value | Source |
|---|---|---|
| Provider | ByteDance (Seed team) | Seed 2.0 launch blog |
| Variants | Pro, Lite, Mini (general agent models); Code (dedicated code model) | Seed 2.0 launch blog |
| API IDs (BytePlus) | seed-2-0-pro-260328, seed-2-0-lite-260428, seed-2-0-mini-260428, seed-2-0-code-preview-260328 | ModelArk model list |
| Model ID (Volcano Engine) | Doubao-Seed-2.0-pro (as named in the model card) | Seed 2.0 model card |
| Release date | 2026-02-14 (launch blog date) | Seed 2.0 launch blog |
| Status | Active on BytePlus; superseded by Seed 2.1 as the newest generation | ModelArk model list; Seed models page |
| Context window | 256K tokens (all four IDs) | ModelArk model list |
| Max input | 224K tokens | ModelArk model list |
| Max output | 128K tokens, including chain of thought; default 4K unless set | ModelArk model list |
| Knowledge cutoff | Not stated by ByteDance on the pages checked (2026-11-15) | Seed 2.0 page; model card |
| Input modalities | Text and image (all IDs); audio (Lite and Mini); video understanding described for the family | ModelArk model list; Seed 2.0 page |
| Output modalities | Text | ModelArk model list |
| Open weights | No | Seed 2.0 page |
| Licence | Proprietary, served through ByteDance channels | Seed 2.0 page; ModelArk model list |
| Parameters | Not stated by ByteDance on the pages checked (2026-11-15) | Seed 2.0 page; model card |
Release and version history
As of 2026-11-15. Dates from ByteDance Seed pages and the BytePlus deprecation notice. The six-digit suffix in each API ID identifies a snapshot; ByteDance does not state that it is a release date, so this page does not treat it as one.
| Date | Event | Source |
|---|---|---|
| 2026-02-14 | Seed 2.0 Pro, Lite, Mini and Code launched; Pro and Code live in the Doubao app and TRAE; full-series API on Volcano Engine | Seed 2.0 launch blog |
| 2026-02-16 | Launch blog cites LMSYS Chatbot Arena standings as of this date | Seed 2.0 launch blog |
| Late April | Lite update adds audio input; Lite described as handling video, image, audio and text | Seed 2.0 page |
| 2026-06-23 | Seed 2.1 announced as the next generation (the blog names Seed2.1 Pro; Turbo appears on the Seed models page) | Seed 2.1 blog |
| 2026-09-08 | First notice of the fourth BytePlus deprecation batch, including early Seed 2.0 Lite and Mini snapshots | ModelArk deprecation notice |
| 2026-09-15 | seed-2-0-mini-260215 and seed-2-0-lite-260228 deprecated (no new endpoints) | ModelArk deprecation notice |
| 2026-11-11 | Those two snapshots deactivated, replaced by seed-2-0-mini-260428 and seed-2-0-lite-260428 (LTS) | ModelArk deprecation notice |
The Seed 2.0 page also links a Pro snapshot named seed-2-0-pro-260328 from its "Try Now" button, matching the BytePlus model list. For dated release history across providers, see the LLM Release Tracker.
How much does Doubao Seed 2.0 cost?
As of 2026-11-15. USD per million tokens, BytePlus ModelArk Standard mode. The tier is set by each request's prompt length. Cache storage is $0.0083 per million tokens per hour for every Standard row. Source: BytePlus ModelArk model pricing page.
| Item | Price | Notes | Source |
|---|---|---|---|
| seed-2-0-pro-260328 | Input $0.50; cached input $0.10; output $3.00 | Prompt 128K to 256K: $1.00 / $0.20 / $6.00. Flex: $0.25 / $0.10 / $1.50 (up to 128K). No Batch price listed | ModelArk model pricing |
| seed-2-0-lite-260428 | Input $0.25; cached input $0.05; output $2.00 | 128K to 256K: $0.50 / $0.10 / $4.00. Flex and Batch: $0.125 / $0.05 / $1.00 (up to 128K). Audio input $3.75, cached $0.75 | ModelArk model pricing |
| seed-2-0-mini-260428 | Input $0.10; cached input $0.02; output $0.40 | 128K to 256K: $0.20 / $0.04 / $0.80. Flex and Batch: $0.05 / $0.02 / $0.20 (up to 128K). Audio input $1.50, cached $0.30 | ModelArk model pricing |
| seed-2-0-code-preview-260328 | Input $0.50; cached input $0.10; output $3.00 | 128K to 256K: $1.00 / $0.20 / $6.00. No Flex or Batch price listed | ModelArk model pricing |
| Model card representative prices | Pro $0.47 / $2.37 (CNY 3.41 / 17.04); Lite $0.09 / $0.53 (CNY 0.64 / 3.83); Mini $0.03 / $0.31 (CNY 0.22 / 2.24) | Input / output per million tokens; the card notes interval-priced models are shown at a single representative price | Seed 2.0 model card |
- Two price sources: the model card table and the BytePlus price list give different figures. Budget against the price list of the platform you actually call; the BytePlus page is the international list.
- Flex mode: charges half the Standard input and output rates, keeps the cached-input price and has no cache-storage fee (BytePlus).
- Free quota: the BytePlus pricing page lists no free quota for Seed 2.0 models.
- Volcano Engine: official CNY price list pages for Volcano Engine were not retrievable on the check date, so only the model card's CNY figures are shown.
Worked example at BytePlus Standard rates: a Seed 2.0 Lite request with 100,000 input tokens and 10,000 output tokens costs $0.025 for input plus $0.02 for output, $0.045 in total. The same request on Pro costs $0.05 plus $0.03, $0.08. For how caching and tiers change the bill, see Model Pricing Explained, and for cross-provider prices the LLM Pricing Tracker.
Context window and output limits
As of 2026-11-15. Source: BytePlus ModelArk model list. Identical values apply to all four Seed 2.0 IDs.
| Limit | Value | Notes | Source |
|---|---|---|---|
| Context window | 256K tokens | Input plus output | ModelArk model list |
| Max input | 224K tokens | Prompts above 128K are billed at the higher tier | ModelArk model list; model pricing |
| Max output | 128K tokens | Includes chain of thought | ModelArk model list |
| Max chain of thought | 128K tokens | Counted within max output | ModelArk model list |
| Default output | 4K tokens | Applies unless max tokens is set | ModelArk model list |
| Rate limits | 30K requests per minute; 1,500K tokens per minute | Stated as theoretical maximums that depend on platform load and invocation method | ModelArk model list |
The 4K default output catches teams out: a long report or a deep reasoning chain is cut off unless the request sets a higher max-tokens value. Because reasoning counts towards the 128K output limit, very long reasoning leaves less room for the visible answer. Context window and output for Volcano Engine endpoints are not stated on the Seed pages checked.
Modalities and tools
The ModelArk model list marks every Seed 2.0 ID as supporting deep reasoning, visual understanding and function calling. Audio understanding is listed for Lite and Mini only. Structured output with a json_schema response format is listed as recommended for Lite and Mini; the list does not mark structured output for Pro or Code.
- Text and images: all four IDs accept text and images and return text.
- Video: the launch blog describes long-form and real-time streaming video understanding, including hour-long videos, and mentions a VideoCut tool.
- Documents: the launch blog describes understanding of complex documents, tables and charts.
- Audio: the Seed 2.0 page says the late-April Lite update made it the series model with unified understanding across video, image, audio and text; BytePlus lists audio input for Lite and Mini, with separate audio prices.
Function calling means the model can request actions from your systems. Treat every tool the model can call as a privileged integration with its own credential and scope. See the LLMjacking Guide.
Where is Doubao Seed 2.0 available?
As of 2026-11-15.
| Channel | Details | Source |
|---|---|---|
| Volcano Engine (China) | Model API for the full Seed 2.0 series; Pro named Doubao-Seed-2.0-pro | Seed 2.0 launch blog; model card |
| BytePlus ModelArk API | Four IDs in ap-southeast-1; Seed 2.0 models also supported in eu-west-1 | ModelArk model list |
| Doubao app | Pro and Code live at launch | Seed 2.0 launch blog |
| TRAE | Pro and Code live at launch | Seed 2.0 launch blog |
| BytePlus Playground | "Try in Playground" link on the Seed 2.0 pages | Seed 2.0 launch blog; Seed 2.0 page |
| Third-party clouds | Not stated by ByteDance on the pages checked (2026-11-15) | Seed 2.0 page |
The BytePlus data processing page states that model deployment and customer data processing take place in Malaysia, Indonesia and/or the EU/EEA. For China transfer rules, see China Cross-Border Data Transfers.
Access restrictions
The Seed 2.0 models are generally available on the channels above. ByteDance does not describe a restricted-access tier, waitlist or trusted-tester programme for Seed 2.0 on the pages checked. Two points limit access in practice:
- Preview ID: the Code model's BytePlus ID carries "preview" in its name (seed-2-0-code-preview-260328) and has no Flex or Batch price.
- Region: seed-2-0-mini-260215 was deactivated in eu-west-1 on 10 September 2026, and the notice directs users to ap-southeast-1. Current snapshots are listed in both regions.
Content in China is also subject to the generative AI rules that apply to public-facing services there; see the China Generative AI Measures and AI Labelling Rules Guide.
Benchmarks (provider-reported, dated)
Provider-reported scores from the Seed 2.0 model card (published around February 2026; checked 2026-11-15). ByteDance chose the benchmarks, settings and comparators; scores are not independently verified here. Independent leaderboard results were not checked for this page.
| Benchmark | Pro | Lite | Mini | Source |
|---|---|---|---|---|
| AIME 2025 | 98.3 | 93.0 | 87.0 | Seed 2.0 model card |
| IMOAnswerBench (no tool) | 89.3 | 81.6 | 71.6 | Seed 2.0 model card |
| Codeforces (Elo) | 3020 | 2233 | 1644 | Seed 2.0 model card |
| LiveCodeBench v6 | 87.8 | 81.7 | 64.1 | Seed 2.0 model card |
| GPQA Diamond | 88.9 | 85.1 | 79.0 | Seed 2.0 model card |
| HLE (no tool, text only) | 32.4 | 28.2 | 13.3 | Seed 2.0 model card |
| SuperGPQA | 68.7 | 67.5 | 61.6 | Seed 2.0 model card |
| MathVista | 89.8 | 89.0 | 85.5 | Seed 2.0 model card |
| MMMU | 85.4 | 83.7 | 79.7 | Seed 2.0 model card |
| VideoMME | 89.5 | 87.7 | 81.2 | Seed 2.0 model card |
| BrowseComp | 77.3 | 72.1 | Not stated by ByteDance | Seed 2.0 model card |
| SWE-Bench Verified | 76.5 | 73.5 | Not stated by ByteDance | Seed 2.0 model card |
The launch blog adds a dated leaderboard reference: as of 16 February 2026 it reports Seed 2.0 at 6th on the LMSYS Chatbot Arena Text Arena (Overall) and 3rd on the Vision Arena. The Seed 2.0 page shows a later table for the updated Lite 0428 snapshot (for example GPQA Diamond 88.4 and BrowseComp 64.0); those figures relate to a different snapshot from the model card. No Code model scores appear in the model card table.
Model card and system card
| Document | What it covers | Source |
|---|---|---|
| Seed 2.0 model card (PDF) | Model IDs, pricing table, capability benchmarks for Pro, Lite and Mini, hallucination metrics (LongFact, FactScore) | Linked from the Seed 2.0 page |
| Seed 2.0 page | Size positioning, Lite audio update, benchmark table for Lite 0428 | seed.bytedance.com/seed2 |
| System card | Not stated by ByteDance on the pages checked (2026-11-15) | Seed 2.0 page; model card |
The model card reports hallucination results as capability metrics rather than as a safety summary, and contains no separate safety evaluation section in the parts read. See Model Cards and System Cards Explained.
Data handling
Source: BytePlus ModelArk data processing page (checked 2026-11-15). These terms apply to the BytePlus international service; Volcano Engine terms were not checked.
- Training use: BytePlus states it does not use customer data for its own model training without the customer's prior authorisation.
- Retention: inputs and outputs that trigger the content filter are retained for 180 days in Malaysia; the service may store customer data for content safety purposes. No other retention period is stated.
- Locations: deployment and processing in Malaysia, Indonesia and/or the EU/EEA.
- Encryption: HTTPS in transit, plus an optional application-layer end-to-end mode that encrypts data to a certificate unique to the inference instance.
- Access control: internal authorisation uses short-term dynamic tokens under a Zero Trust Identity system.
The model card mentions a Data Authorisation Agreement for an incentive programme that uses authorised usage data; read it before opting in. For cross-provider comparison, see LLM Data Retention and Training-Use Policies by Provider.
Deprecation status
As of 2026-11-15. Source: BytePlus ModelArk model deprecation notice and model list.
| Model or item | Details | Source |
|---|---|---|
| seed-2-0-pro-260328 | Active; no deprecation notice listed | Deprecation notice; model list |
| seed-2-0-lite-260428 | Active; named as LTS replacement | Deprecation notice |
| seed-2-0-mini-260428 | Active; named as LTS replacement | Deprecation notice |
| seed-2-0-code-preview-260328 | Active; no deprecation notice listed | Deprecation notice; model list |
| seed-2-0-lite-260228 | Deprecated 15 September 2026; deactivated 11 November 2026; automatic replacement seed-2-0-lite-260428 | Deprecation notice |
| seed-2-0-mini-260215 | Deprecated 15 September 2026; deactivated 11 November 2026 (10 September 2026 in eu-west-1); automatic replacement seed-2-0-mini-260428 | Deprecation notice |
| Notice policy | At least 30 calendar days before deactivation, via in-site messages and platform announcements; no new endpoints or fine-tuning after the deprecation date; quotas reduced from first notice | Deprecation notice |
Because the snapshot suffix changes with each update, pin the full dated ID in production and treat an automatic replacement as a model change that needs regression testing.
How Doubao Seed 2.0 differs from other models in its family
| Model | Positioning (provider wording) | Context / max output | Standard input / output (USD, up to 128K) | Audio input |
|---|---|---|---|---|
| Seed 2.0 Pro | Long-chain reasoning and robustness in complex workflows | 256K / 128K | $0.50 / $3.00 | No |
| Seed 2.0 Lite | Balances output quality and response speed | 256K / 128K | $0.25 / $2.00 | Yes |
| Seed 2.0 Mini | Inference throughput and deployment density | 256K / 128K | $0.10 / $0.40 | Yes |
| Seed 2.0 Code (preview) | Dedicated code model | 256K / 128K | $0.50 / $3.00 | No |
All four share the same limits and capability flags apart from audio and structured output, so the choice between them is mainly cost, throughput and the provider-reported score gap on your task type. Seed 2.1 is the next generation: the Seed models page lists Pro and Turbo sizes, with the BytePlus ID dola-seed-2-1-turbo-260628. Its details sit on the ByteDance Doubao Seed Models page. For cross-provider positioning, see Frontier AI Models Compared and DeepSeek vs Qwen vs Kimi vs GLM.
Practitioner checklist
- Pin the full dated ID (for example seed-2-0-lite-260428) and subscribe to BytePlus deprecation notices.
- Set max tokens explicitly; the 4K default output truncates long answers and reasoning.
- Budget against the price list of the platform you call, using the 128K prompt threshold and reasoning tokens.
- Choose ap-southeast-1 or eu-west-1 deliberately and record the processing locations in your data transfer assessment.
- Confirm whether you have authorised any training use, and do not opt into data programmes without review.
- Use json_schema structured output on Lite and Mini where downstream code parses responses.
- Scope every function-calling tool to least privilege and keep API keys in a secrets manager with rotation.
- Regression-test on each automatic snapshot replacement before it reaches production traffic.
Standards and references
- ByteDance Seed: Seed 2.0 Official Launch (accessed 2026-11-15)
- ByteDance Seed: Seed 2.0 page (accessed 2026-11-15)
- ByteDance Seed: Seed 2.0 Model Card (PDF) (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)
- ByteDance Seed: Models (accessed 2026-11-15)
- ByteDance Seed: Seed2.1 Officially Released (accessed 2026-11-15)
Related NHI Mgmt Group resources: ByteDance Doubao Seed Models, LLM Pricing Tracker, Frontier AI Models Compared, LLM Data Retention and Training-Use Policies by Provider, LLMjacking Guide.