Seed 2.1 (Doubao)

ByteDance's flagship Seed 2.1: Pro for deep-thinking agent work, Turbo for scale, plus the weekly-updated Doubao-Seed-Evolving. Closed weights.

Last reviewed September 6, 2026

At a glance

Overview

Seed 2.1 is ByteDance’s current flagship model family, released on June 23, 2026 at the Volcano Engine FORCE conference and still unsuperseded as of September 2026. ByteDance Seed’s model directory describes it in a line: “A next-generation agent for real-world productivity.” Where Seed 2.0 was pitched as a broad multimodal upgrade, Seed 2.1 is pitched as an agent — measured by whether it carries a real task from requirement to validated result. The launch post leads with “more reliable general agent capabilities” for multi-step workflows and “more stable end-to-end coding delivery.”

The family ships as two SKUs plus one moving target:

At launch ByteDance claimed three capabilities — coding engineering delivery, agent long-chain task execution, and multimodal understanding — were “comparable to GPT-5.5.” Read that as a dated statement: it was made on June 23, 2026, against a model OpenAI has since replaced twice, most recently with GPT-6 Astra on September 4, 2026. It is evidence about Seed 2.1’s June position, not about where it sits against today’s frontier.

Seed 2.1 arrived alongside other FORCE announcements — Seedance 2.5 for video, an audio model, and Seedream 5.0 Pro for images. A preview had surfaced four days earlier; ByteDance’s post “Seed-2.1-Preview Model Release on Arena” names the build it published to Arena AI’s Code arena as Seed-2.1-Pro-Preview.

Model Lineage

ByteDance’s naming has been widely misreported, so the lineage is worth stating precisely.

Seed 2.0 launched on February 14, 2026: three general-purpose agent models — Pro, Lite, and Mini — “along with a dedicated Code model.” Seed 2.0 Lite was upgraded at the end of April 2026 into the series’ first omni-modal understanding model, unifying video, image, audio, and text input. There was no Vision variant in Seed 2.0, and there is none in Seed 2.1 — vision is native to the standard models.

Seed 2.1 ships as exactly two SKUs; ByteDance Seed’s product page is explicit that the family offers “two AI productivity models in different sizes: Pro and Turbo.” There is no plain “Seed 2.1” endpoint, and no Seed 2.1 Code variant — Doubao Seed 2.0 Code remains ByteDance’s dedicated coding model.

Doubao-Seed-Evolving, added to the Volcano Engine Coding Plan on August 21, 2026, is less a third SKU than a different release model: a single ID that ByteDance re-points at improved weights at least once a week, with no dated snapshot underneath. Reporting describes it as currently equivalent in capability to Seed 2.1 Pro, extended to a 1M context window.

Capabilities

• General agent execution: More reliable multi-step workflow handling — project planning, document processing — which ByteDance positions as the headline improvement over Seed 2.0.
• Long-horizon planning and repair (Pro): The Pro tier is sold on requirement understanding, planning across long chains, and continuous repair — recovering from its own failed steps rather than ending the run.
• End-to-end coding delivery: The full loop, from requirement analysis through implementation to result validation, rather than isolated code completion.
• Mobile GUI operation: Reads screen content, decides the next action, and executes sequential taps, typing, and app switches.
• Multimodal understanding: Native handling of images, charts, documents, and video, which ByteDance describes as further enhanced over Seed 2.0.
• Frontend generation: Top-10 placement in 5 of 7 frontend subcategories on Code Arena.
• Long context on Evolving: 1M input tokens, four times the window of the dated SKUs.

Technical Specifications

ByteDance publishes little about how Seed 2.1 is built, but it does publish the operational limits. Only vendor-stated figures appear below; the Chinese limits come from Volcano Engine’s model list.

• Ark model IDs: doubao-seed-2-1-pro-260628, doubao-seed-2-1-turbo-260628, doubao-seed-evolving (plus the Seed-2.1-Pro-Preview build published to Arena)
• Modalities: Text, image, and video input; text output. Audio input is supported in the Seed 2.0 Lite omni-modal tier.
• Context window: 256K tokens on Pro and Turbo (上下文窗口 256k), identical for both; 1M tokens on Doubao-Seed-Evolving.
• Max output: 256K tokens, default 4K. Max input and max chain-of-thought: 256K on Pro and Turbo.
• Rate limits: 500 RPM, 1,000,000 TPM — flagged by Volcano Engine as 非刚性保障 (“not a rigid guarantee”), varying with platform load.
• Versioning: Pro and Turbo are pinned by date suffix and support batch inference. Evolving is not versioned at all — one ID, at least one upgrade per week, no snapshots, no batch.
• Availability: Doubao app, Volcano Engine (Ark) API, and the Ark Coding Plan subscription.
• Weights: Closed. No open-weight release.
• Parameter count, knowledge cutoff, architecture: none published. ByteDance has not confirmed whether Seed 2.1 uses a mixture-of-experts design, despite widespread assumptions that it does.

Use Cases

The three SKUs are not interchangeable, and the routing decision is mostly about how much you are willing to pay for planning depth and how much version stability you need.

Route to Seed 2.1 Pro when the task is hard and the run is long:

• Multi-step engineering delivery — requirement to implementation to validated result. This is the workload ByteDance optimized Pro for, and where its Agent Startup Bench lead over Turbo (68.8 vs 54.0) is widest.
• Long-chain agents that hold a plan across dozens of tool calls and recover from their own failures.
• Complex coding where requirement interpretation, not code completion, is the bottleneck.
• Document, chart and video understanding — the areas ByteDance highlights via CharXiv-RQ, MeasureBench, TVBench, and TOMATO.

Route to Seed 2.1 Turbo when volume and latency dominate:

• High-concurrency production serving at half the token price — the tier’s entire reason to exist.
• Consumer assistant workloads: Seed powers the Doubao app, so the line is tuned for latency-sensitive serving at volume.
• Mobile automation and GUI agents, where Turbo’s MobileWorld score (70.0) lands within 3.1 points of Pro’s at half the price. This is the one benchmark family where the “comparable to Pro” pitch holds up.

Route to Doubao-Seed-Evolving only when freshness beats reproducibility:

• Interactive coding and agent work inside the Coding Plan, where a human is in the loop and better weights next week are a straight win.
• Anything needing more than 256K of context — it is the only Seed SKU with a 1M window.
• Not for evaluation baselines, regression suites, batch jobs, or any output you may have to defend later.

Performance / Benchmarks

Everything below is vendor-run by ByteDance and dated to the June 2026 launch. The comparison set — GPT-5.5, Claude-4.7-Opus, Gemini-3.1-Pro — has since been replaced by newer models at all three labs, so these tables establish where Seed 2.1 stood in June, not where it stands now. The launch post states most results as relative claims; the model cards publish absolute scores, and we use those. ByteDance publishes no independent aggregate-index score for the family.

Seed 2.1 (Seed2.1 Model Card, June 2026)

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:

Note what the Turbo column does to the “comparable to Pro” pitch. On GUI work it holds — MobileWorld 70.0 against 73.1. On agentic delivery it does not: Agent Startup Bench 54.0 against 68.8, a 14.8-point gap, and CreativeWork 34.5 against 42.5. Turbo is close on some axes, not on all of them.

The model card contradicts itself. Its prose gives Seed2.1-Pro 73.3 on “MobileWorld GUI-only” and OSWorld 78.2 where Table 4 says 73.1 and 78.8, and quotes baselines that match neither. We report the table values and flag the discrepancy rather than silently picking one.

The launch post adds four claims without absolute scores: Seed 2.1 Pro “achieves the highest scores” on CharXiv-RQ and MeasureBench (visual understanding), “scores at industry-leading levels” on TVBench and TOMATO (video), “ranks among the top tier” on Agents’ Last Exam (model card: 19.5 / 41.4), and shows “standout performance” on MMLongBench-128K.

Doubao-Seed-Evolving has no published benchmark results at all, and by construction cannot have durable ones: any score measures one week’s weights.

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

On Code Arena: Frontend, ByteDance reports a rank of 8th with a score of 1539 and “a top-10 position in 5 out of 7 frontend subcategories.” It credits “the Seed2.1 Preview,” but the build it published to the arena was Seed-2.1-Pro-Preview — so the result belongs to the Pro preview, not to Turbo or the family. It is also the Frontend board, not an overall ranking, and a June 2026 placement on a board that moves daily.

Seed 2.0 Pro (Seed2.0 Model Card, February 2026)

Kept because readers still arrive here from old links. Four widely-circulated “Seed 2.0” figures are genuine Seed2.0 Pro results from ByteDance’s own Seed2.0 Model Card: AIME 2025 98.3 (GPT-5.2 High 99.0), Codeforces Elo 3020 without tools (GPT-5.2 High 3148), SWE-Bench Verified 76.5, and VideoMME 89.5 with subtitles (Gemini-3-Pro 88.4). Two cautions: SWE-Bench 76.5 is not a leadership claim, trailing GPT-5.2 High and both Claude entrants; and the VideoMME table fields no GPT or Claude entrant at all, so 89.5 is a win over a narrower field.

Limitations

• Doubao-Seed-Evolving cannot be pinned, which makes it unfit for anything auditable. One ID, at least one weights change per week, no dated snapshots, and no batch inference. A regression suite that passes on Monday can fail the following Monday with no change on your side and no version number to diff against; a published evaluation number describes a model that no longer exists. If your output has to be defensible later, call doubao-seed-2-1-pro-260628 or doubao-seed-2-1-turbo-260628 and accept the older weights. Third-party gateways make the point involuntarily: several still advertise Evolving with the 256K window it launched with rather than the 1M it now has.
• No open weights. Seed 2.1 is closed and API-only, and the flagship Doubao line has stayed closed through August 2026 — ByteDance releases smaller open Seed-OSS models, but nothing at this tier. Self-hosting, on-premise fine-tuning, and offline deployment are all off the table.
• Little is published about the architecture. No parameter count, no knowledge cutoff, no confirmed architecture family — the operational limits are published, the internals are not.
• Benchmark claims are vendor-run, dated, and internally inconsistent. The Seed2.1 model card disagrees with itself on MobileWorld (73.1 vs 73.3) and OSWorld (78.8 vs 78.2), its comparison set is June 2026, and its “comparable to GPT-5.5” line is now two OpenAI generations behind. Independent replication is scarce.
• “Comparable to Pro” oversells Turbo on agentic work — 14.8 points behind on Agent Startup Bench, in ByteDance’s own table. Benchmark both on your workload before routing production traffic to the cheaper tier.
• China-first availability, and CNY-only pricing. Documentation is far richer in Chinese than in English, and Seed 2.1 had not appeared on BytePlus ModelArk, the international storefront, as of late August 2026. There is no ByteDance USD rate card, so any dollar figure is a third-party conversion at a drifting rate — the published ones already disagree by about 6%.
• Content policy 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 invents variants ByteDance never shipped (a “Seed 2.0 Vision,” a “Seed 2.1 Code”) and often drops Turbo or Evolving from the family. Verify model IDs against Volcano Engine before building against them.

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, and nothing below is converted. Under 在线推理（常规）(“online inference, standard”) the SKUs bill at a single input-length tier, 输入长度 [0, 256], so there is no length-based price step of the kind Seed 2.0 had:

Input, cache-hit and output rates are 元/百万token (yuan per million tokens); cache storage is 元/百万token/小时. All figures are CNY, reproduced unconverted; the audio-input columns are empty (-) for every SKU. 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, ¥1.50 / ¥7.50 for Turbo — with cache-hit rates unchanged and no cache-storage charge. Evolving is excluded from batch entirely.

The one USD figure ByteDance has ever published is for the previous generation: the Seed2.0 Model Card’s “Token Prefill / Decode Price Comparison (USD per 1M tokens)” lists Seed2.0 Pro at $0.47 input / $2.37 output, and says it reports “a single representative price” for models with interval pricing. It is illustrative, and it does not describe Seed 2.1.

Access runs through four surfaces: the Doubao app (“Doubao and Volcano Engine users can now start to access Doubao Seed 2.1”), the Volcano Engine (Ark) API — the only general programmatic route — the Ark Coding Plan, a flat-rate developer subscription whose model pool is revised through the year and which Doubao-Seed-Evolving entered on August 21, 2026, and TRAE, ByteDance’s AI IDE. Rates change; confirm against the 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.
• Seedream 5.0 and Seedance 2.5 — the image and video generation lines, both served through Volcano Ark.
• Seed1.5-VL — the only dedicated vision-language SKU in the catalogue.

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) — GDPval, MobileWorld, Agent Startup Bench scores
• Seed2.0 Model Card (PDF) — the AIME, Codeforces, SWE-Bench and VideoMME figures
• Seed 2.0 Official Launch — 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