Doubao Seed Evolving

What it is. Doubao Seed Evolving is ByteDance’s rolling, versionless model, first released June 2026 and updated at least once per week under a single stable model ID. Why it matters. It is ByteDance’s continuously-updated 1M-context model, positioned for agent and coding scenarios, and it represents a distinct delivery model — “no versions” — that trades reproducibility for always-current capability. Key characteristics. 1M-token context (raised from 256K in a July 2026 update), 262,144 max output, deep thinking, multimodal understanding, GUI task handling, tool calling and structured output. What a professional should know. Capabilities can change under the same model ID, architecture is not disclosed, the API is cn-beijing only, and pricing is 6.00 CNY input / 30.00 CNY output per 1M tokens (cache hits 1.20).

Architecture and parameters

ByteDance discloses no architecture or parameter counts for the Seed series, and the database records that absence. China AI Hub analysis indicates the rolling model compounds the opacity: not only is the architecture hidden, but the very identity of the model changes weekly, so any performance characterization is a moving target. See Mixture-of-Experts for the architectural context ByteDance does not provide.

What the context window actually means

The 1M-token context (raised from 256K in July 2026) with a 262,144 output ceiling positions Seed Evolving for the long-horizon agent and deep-research workloads ByteDance names, with a mid-tier generation headroom. See the context-window research and the long-context hub.

Pricing implications

6.00 CNY input / 30.00 CNY output per 1M tokens, cache hits 1.20 (20% of input) — identical list price to the Seed 2.1 Pro model. China AI Hub analysis indicates the 5x input-to-output ratio is the Seed family’s generation-premium signature, and Evolving’s price parity with Pro means the premium is for the rolling update cadence and 1M context, not for a higher capability tier. See choosing-by-price.

API, coding, and agent implications

ByteDance positions Seed Evolving for agent and coding scenarios, with tool calling, structured output, GUI task handling and computer use documented. The operational caveat is the rolling model ID: capabilities can change week to week, so any pinned integration or evaluation is at risk of drift. See computer use and AI agents.

Open weights and license

Proprietary, closed weight, no self-hosting. See open weight vs API.

Benchmark interpretation

No benchmark scores are recorded for Seed Evolving in the database. This is an absence of evidence, and the rolling nature makes it structurally hard to benchmark — any published number would describe only the snapshot at that moment.

What the benchmarks do not prove

With no recorded benchmarks, there is nothing with which to rank Seed Evolving against competitors, and its agent/coding positioning rests on vendor claims rather than published numbers. See how to read vendor-reported benchmarks.

Suitable and less suitable workloads

Well-suited: agent and coding workloads inside ByteDance’s China ecosystem that benefit from weekly capability updates and 1M context, and teams willing to accept version drift for always-current output. Less suited: reproducibility-sensitive production (no pinned version), international API consumers (cn-beijing only), and any adoption decision requiring published architecture or benchmark evidence.

Where this model fits

Workload	Relevance
Long-context analysis	High
Coding	Moderate
Structured API workflows	High
Agent orchestration	High
Local self-hosted deployment	No evidence
GUI automation	High
Video generation	No evidence
Enterprise cloud	Moderate

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

China AI Hub analysis

China AI Hub analysis indicates Seed Evolving is ByteDance’s bet on continuous delivery as a differentiator: a versionless, weekly-updated 1M model that competes on always-current capability rather than on a documented checkpoint. Against the versioned Seed 2.1 Turbo (256K, half price) and Seed 2.1 Pro (1M, same price), Evolving sits at the top on context and update cadence but sacrifices the reproducibility that a pinned version provides — a genuine trade-off, not a pure upgrade. The combination of rolling identity, hidden architecture and cn-beijing-only serving makes independent verification unusually difficult, which is the defining constraint for international adopters. See the state of China’s AI models and the rise of Chinese AI agents research.

What is uncertain

• Architecture and parameter counts are not publicly disclosed.
• No benchmark scores are published, and the rolling model makes any score snapshot-specific.
• Capabilities can change weekly under the same model ID.
• The API is served from cn-beijing only; no international endpoint was verified.

Sources

• Ark official model list
• Ark official model pricing
• Ark model release announcements

Labels used above: Official fact (context, capabilities, pricing and the rolling update model from Ark’s docs) and China AI Hub analysis (our synthesis, always introduced as such). No benchmark evidence is currently recorded for this model.