Seed 2.1 (Doubao)

Overview

Seed 2.1 is ByteDance's current flagship model family, released on June 23, 2026 at the Volcano Engine FORCE conference. ByteDance Seed's own model directory describes it in a single line: "A next-generation agent for real-world productivity."

That framing is the point. Where Seed 2.0 was pitched as a broad multimodal upgrade, Seed 2.1 is pitched as an agent — a model measured by whether it can carry a real task from requirement to validated result. ByteDance's launch post leads with "more reliable general agent capabilities" for multi-step workflows such as project planning and document processing, and "more stable end-to-end coding delivery" spanning requirement analysis through result validation.

Seed 2.1 arrived alongside a cluster of other FORCE announcements, including Seedance 2.5 for video, an audio generation model, and Seedream 5.0 Pro for images — ByteDance assembling a full multimodal production stack under the Volcano Engine umbrella.

A preview build surfaced publicly a few days before the formal launch, on June 19, 2026. ByteDance's post — titled "Seed-2.1-Preview Model Release on Arena" — states that the model it released to Arena AI's Code arena was Seed-2.1-Pro-Preview, and that it would reach Feishu Spark and Coze "within the next two weeks."

Model Lineage

Understanding Seed 2.1 requires understanding what shipped before it, because ByteDance's naming has been widely misreported.

Seed 2.0 launched on February 14, 2026. ByteDance's launch post enumerates the family exhaustively: three general-purpose agent models — Pro, Lite, and Mini — "along with a dedicated Code model." On the ByteDance Seed site the three general models are differentiated as:

There was no Vision variant in the Seed 2.0 family, and there is none in Seed 2.1. Vision is native to the standard models.

Seed 2.0 Lite was upgraded at the end of April 2026 to become the Seed series' first omni-modal understanding model, natively unifying video, image, audio, and text input in a single model, with upgraded agent, coding, and GUI capabilities alongside.

Seed 2.1, released June 23, 2026, sits at the top of the lineup and ships as exactly two SKUs. ByteDance Seed's product page is explicit: "The Seed2.1 model family is officially released, offering two AI productivity models in different sizes: Pro and Turbo." There is no plain "Seed 2.1" endpoint — on Volcano Engine the two model IDs are doubao-seed-2-1-pro-260628 and doubao-seed-2-1-turbo-260628.

Seed 2.1 did not replace Doubao Seed 2.0 Code, which remains ByteDance's dedicated coding model — no Seed 2.1 Code variant exists.

Capabilities

• General agent execution: More reliable multi-step workflow handling, including project planning and document processing. ByteDance positions this as the headline improvement over Seed 2.0.
• End-to-end coding delivery: Support across the full loop — requirement analysis, implementation, and result validation — rather than isolated code completion.
• Mobile GUI operation: The model reads screen content, decides the next action, and executes sequential operations such as tapping, typing, and switching between apps.
• Multimodal understanding: Native handling of images, charts, documents, and video. ByteDance says the Seed 2.0 generation "can now parse complex documents, tables, graphs, and video content with significantly improved accuracy," and describes Seed 2.1's multimodal and foundational reasoning as further enhanced.
• Frontend generation: Strong showing on frontend-focused code evaluation, with top-10 placement in 5 of 7 frontend subcategories on Code Arena.
• Omni-modal input (Lite tier): Seed 2.0 Lite adds audio to video, image, and text, enabling cross-modal reasoning that fuses speech and ambient sound with visual signals.

Technical Specifications

ByteDance publishes little about how Seed 2.1 is built, but it does publish the operational limits. We list only what the vendor has stated.

• Variants: Seed 2.1 Pro and Seed 2.1 Turbo — the whole family (plus the Seed-2.1-Pro-Preview build published to Arena)
• Ark model IDs: doubao-seed-2-1-pro-260628, doubao-seed-2-1-turbo-260628
• Modalities: Text, image, and video input; text output. Audio input is supported in the Seed 2.0 Lite omni-modal tier.
• Availability: Doubao app and the Volcano Engine (Ark) API platform
• Weights: Closed. No open-weight release.
• Context window: 256K tokens, identical for both SKUs.
• Max input: 256K tokens. Max output: 256K tokens (default 4K). Max chain-of-thought: 256K tokens.
• Rate limits: 500 RPM, 1,000,000 TPM — which Volcano Engine explicitly flags as 非刚性保障 ("not a rigid guarantee"), varying with platform load.
• Parameter count: Not published by ByteDance.
• Architecture: Not disclosed. ByteDance has not confirmed whether Seed 2.1 uses a mixture-of-experts design, despite widespread assumptions that it does.
• Knowledge cutoff: Not published by ByteDance.

Volcano Engine's model list states the length limits in Chinese as 上下文窗口: 256k / 最大输入: 256k / 最大回答(默认 4k): 256k / 最大思维链: 256k — "context window: 256k / max input: 256k / max output (default 4k): 256k / max chain-of-thought: 256k."

Use Cases

• Multi-step office workflows: Project planning, document processing, and other tasks that require holding a plan across many turns.
• End-to-end software delivery: Taking a requirement, producing an implementation, and validating the result — the workflow ByteDance explicitly optimized for.
• Mobile automation and GUI agents: Driving phone interfaces through sequential tap, type, and app-switch operations.
• Document and chart intelligence: Parsing complex documents, tables, and graphs where layout carries meaning.
• Video comprehension: Understanding temporal content, which ByteDance highlights via TVBench and TOMATO results.
• Consumer assistant workloads: Seed powers the Doubao app, one of China's most-used AI applications, so the model is tuned for high-volume, latency-sensitive consumer serving.

Performance / Benchmarks

Everything below is vendor-reported by ByteDance, either from its own evaluation or from its reading of third-party leaderboards. ByteDance's launch post states most results as relative claims, but its model cards publish absolute scores, and we use those in preference to the prose.

Seed 2.1 (Seed2.1 Model Card)

Table 1 of the model card compares GPT-5.5, Claude-4.7-Opus, Gemini-3.1-Pro, Seed2.1-Turbo, and Seed2.1-Pro:

Table 4, on generalist computer use:

The Seed2.1 Model Card contradicts itself on MobileWorld. Table 4 gives Seed2.1-Pro 73.1; the prose of the same document says Seed2.1-Pro reaches 73.3 "on MobileWorld GUI-only," and quotes baselines (GPT-5.5 62.4, Claude-4.7-Opus 56.4, Gemini-3.1-Pro 58.1) that do not match its own table either. The prose likewise reports OSWorld 78.2 where the table says 78.8. We report the table values and flag the discrepancy rather than silently picking one.

From the launch post, stated without absolute scores:

• CharXiv-RQ and MeasureBench — Seed 2.1 Pro "achieves the highest scores," for visual understanding.
• TVBench and TOMATO — Seed 2.1 Pro "scores at industry-leading levels," for video comprehension.
• Agents' Last Exam (ALE) — Seed 2.1 Pro "ranks among the top tier of participating models." (Model card: 19.5 / 41.4.)
• MMLongBench-128K — "standout performance" on long context.

Code Arena (third-party leaderboard, as reported by ByteDance)

• Code Arena: Frontend — ranks 8th with a score of 1539, and "secures a top-10 position in 5 out of 7 frontend subcategories."

The launch post credits this to "the Seed2.1 Preview." The build ByteDance actually published to Arena AI's Code arena was Seed-2.1-Pro-Preview, per its own Arena announcement — so the result belongs to the preview of the Pro tier, not to the family as a whole, and not to Turbo. Note this is the Frontend leaderboard, not an overall ranking.

LMArena (third-party leaderboard, as reported by ByteDance, for Seed 2.0)

As of February 16, 2026, ByteDance reported that Seed ranked 6th on the LMArena Text Arena (Overall) leaderboard and 3rd on the Vision Arena leaderboard. These placements describe Seed 2.0, not Seed 2.1, and leaderboards move continuously — treat them as a February 2026 snapshot.

Seed 2.0 Pro (Seed2.0 Model Card)

Four figures that circulate widely for "Seed 2.0" are genuine, and they are Seed2.0 Pro results published in ByteDance's own Seed2.0 Model Card. They describe the previous generation, not Seed 2.1:

The model card's prose corroborates two of these directly: Seed2.0 Pro "reaches a Codeforces Elo of 3020," and "achieves a breakthrough performance of 89.5 on VideoMME."

Two cautions. On SWE-Bench Verified, 76.5 is not a leadership claim — it trails GPT-5.2 High (80.0), Claude-Opus-4.5 (80.9) and Claude-Sonnet-4.5 (77.2), and edges out only Gemini-3-pro High (76.2). And the card says nothing about Doubao Seed 2.0 Code, which it does not benchmark at all. The VideoMME table compares Seed2.0 against Gemini-3-Pro, Gemini-3-Flash, and earlier Seed models only — no GPT or Claude entrant — so 89.5 is a win over a narrower field than the other rows.

Limitations

• Little is published about the architecture. No parameter count, no knowledge cutoff, no confirmed architecture family. Context window and rate limits are published; the internals are not.
• Benchmark claims are vendor-reported, and the model card is not self-consistent. Seed2.1's own model card disagrees with itself on MobileWorld (73.1 vs 73.3) and OSWorld (78.8 vs 78.2). Independent replication is scarce.
• No open weights. Self-hosting, fine-tuning on your own infrastructure, and offline deployment are all off the table.
• China-first availability. The primary surfaces are the Doubao app and Volcano Engine, both oriented toward the Chinese market. Documentation is substantially richer in Chinese than in English.
• Content policy and regulatory constraints. As a Chinese-market model, Seed operates under content rules that differ from those governing US-hosted models, and outputs may reflect this.
• Naming confusion in the wild. Secondary coverage frequently invents variants (a "Seed 2.0 Vision," a "Seed 2.1 Code") that ByteDance never shipped, and often drops Turbo from the Seed 2.1 family. Verify model IDs against Volcano Engine before building against them.
• Pricing is published in CNY, not USD. The Seed 2.1 rate card is denominated in yuan and hosted on Volcano Engine's Chinese-language pricing page. ByteDance has published USD figures only for the older Seed 2.0 generation.

Pricing & Access

ByteDance publishes a Seed 2.1 rate card, in CNY, on Volcano Engine's 模型价格 (model pricing) page. It is Chinese-language only; there is no USD rate card for Seed 2.1.

Under 在线推理（常规）— "online inference (standard)" — both SKUs are billed at a single input-length tier, 输入长度 [0, 256] (input length 0–256 thousand tokens), so there is no length-based price step the way there is for Seed 2.0:

Input, cache-hit and output rates are 元/百万token (yuan per million tokens); cache storage is 元/百万token/小时 (yuan per million tokens per hour). The audio-input columns are empty (-) for both SKUs.

Under 批量推理 ("batch inference"), the published rates are exactly half the standard input and output rates — ¥3.00 / ¥15.00 for Pro and ¥1.50 / ¥7.50 for Turbo — while cache-hit rates are unchanged. Batch has no cache-storage charge.

The one USD figure ByteDance itself publishes is for the previous generation. The Seed2.0 Model Card gives a "Token Prefill / Decode Price Comparison (USD per 1M tokens)" listing Seed2.0 Pro at $0.47 input / $2.37 output, with the caveat that "for Seed2.0 models with interval pricing, we report a single representative price" — so it is illustrative, not a rate card, and it does not describe Seed 2.1.

Access channels:

• Doubao app — the consumer assistant. ByteDance states that "Doubao and Volcano Engine users can now start to access Doubao Seed 2.1."
• Volcano Engine (Ark) — the API and model platform. This is the only programmatic route.
• TRAE — ByteDance's AI IDE. Seed 2.0 Pro and Seed 2.0 Code launched on the Doubao app and TRAE together.

Rates change; confirm against the Volcano Engine pricing page before budgeting.

Ecosystem & Tools

• Volcano Engine (Ark) — ByteDance's cloud AI platform and the API surface for every Seed model.
• Doubao app — the consumer assistant that Seed powers.
• TRAE — ByteDance's AI-native IDE, where Doubao Seed 2.0 Code runs in both SOLO and IDE modes.
• Seedream 5.0 — the image generation line, also served through Volcano Ark.
• Seed1.5-VL — ByteDance Seed's standalone vision-language model, and the only dedicated VL SKU in the catalogue.
• Seedance — the video generation line; Seedance 2.5 was previewed at the same FORCE conference that launched Seed 2.1.

Community & Resources

• Seed 2.1 Officially Released: Advancing AI Productivity — the official launch post
• Seed2.1 product page — source of the "Pro and Turbo" statement
• Seed2.1 Model Card (PDF) — absolute scores for GDPval, MobileWorld, Agent Startup Bench
• Seed2.0 Model Card (PDF) — source of the AIME 98.3, Codeforces 3020, SWE-Bench 76.5 and VideoMME 89.5 figures
• Seed 2.0 Official Launch — enumerates the Pro / Lite / Mini / Code family
• Seed-2.1-Preview Model Release on Arena — the June 19, 2026 preview; names the build Seed-2.1-Pro-Preview
• Volcano Engine model list — Ark model IDs and the 256K context window
• Volcano Engine model pricing — the CNY rate card
• ByteDance Seed Models directory — canonical list of every Seed model
• ByteDance Seed — research team homepage