Overview
Doubao Seed Evolving is a model line on Volcengine Ark, called through the single model ID doubao-seed-evolving. New versions take effect server-side, so developers never migrate an endpoint or change request parameters. It targets agent and coding workloads, launched on June 23, 2026, moved to a 1M context on July 14 and has iterated on a rolling cadence since. The latest update on November 5 focused on task planning and result verification for multi-step work plus tool calling and factual consistency, and ByteDance says response latency and per-call cost dropped across most scenarios.
Key Features
- One model ID that keeps upgrading: New versions roll out server-side while the calling convention stays fixed, which saves endpoint migrations and regression runs, at the cost of not being able to pin one exact version for long-term reproducibility.
- 1M context and 256K output: A 1024K context window, up to 1024K input and 256K output, with the chain of thought also capped at 256K, so long documents and repository-scale code fit in one request.
- Trained for agent work: Aimed at multi-step, long-horizon tasks, the model breaks down goals, keeps advancing and judges when the work is actually complete, cutting the back-and-forth confirmation loop.
- Anti-hallucination and factual consistency: When information is uncertain it retrieves and verifies before answering and marks the scope of its conclusions; the November 5 update specifically targeted tool calling and factual consistency.
- Full multimodal understanding: Document, image and video understanding arrive alongside structured output and context caching, which helps turn unstructured material into something computable first.
- Transparent rate limits and pricing: Up to 500 RPM and 1,000,000 TPM; input 6 RMB and output 30 RMB per million tokens, cache hits at 1.2 RMB, cache storage at 0.017 RMB per million tokens per hour.
Use Cases
- Coding agents: repository-scale refactors, multi-turn tool calls and test backfilling, with the 1M context removing a separate retrieval step.
- Long-horizon business tasks: support, sales and research summarization that need many consecutive steps and a clear definition of done.
- Document processing: structured extraction from contracts, drawings and reports, writing straight into a database through structured output.
- Drawing and 3D understanding: read engineering drawings and technical documents, then recommend the next action.
- High-concurrency traffic: the 500 RPM allowance makes it a reasonable default model for batch workloads.
Pros
- One model ID always points at the newest version, so integration happens once.
- 1M context plus 256K output removes the need to slice long documents and repository code first.
- Training emphasis sits on agents and coding, the two areas ByteDance keeps highlighting for completion and stability.
- The November 5 release deliberately strengthened anti-hallucination and factual consistency, exactly where long tasks usually break.
- OpenAI-compatible endpoints, so the Ark SDK and common clients work without adapters.
- Cache hits at 1.2 RMB per million tokens bring real bills well under list price for long multi-turn work.
Pricing
Pricing is flat across 0 to 1024K input length: 6 RMB per million input tokens, 30 RMB per million output tokens, 1.2 RMB for cache hits and 0.017 RMB per million tokens per hour for cache storage. Volcengine Ark also sells Agent and Coding Plan subscriptions. The TRAE enterprise edition bundles the same model at matching list prices.
Summary
Seed Evolving does not sell a version number, it sells a model ID that keeps tracking the newest capability, so integration happens once. The 1M context and 256K output remove the pre-slicing step for long work, and the November 5 release tightened anti-hallucination and tool-calling stability. Teams that must pin a version for reproducibility need their own snapshots; everyone else gets a low-maintenance default.
Version History
- Seed Evolving rolls out an update focused on agent execution and anti-hallucination (2026-11-05): On November 5 Volcengine Ark updated doubao-seed-evolving: task planning and result verification for multi-step work became more complete, tool calling and factual consistency improved noticeably, and the model now retrieves and verifies before answering when information is uncertain, while response latency and per-call cost dropped in most scenarios. The upgrade takes effect server-side and calls still go through the single model ID. Current specs are a 1024K context, 256K maximum output, a 256K chain of thought, 500 RPM and 1,000,000 TPM, at 6 RMB input and 30 RMB output per million tokens.
- Doubao Seed Evolving launches, context reaches 1M in July (2026-06-23): The Doubao Seed family added Seed Evolving on Volcengine Ark, aimed at agent and coding work and delivered through a single model ID with rolling upgrades. On July 14 it gained a 1M context plus better long-horizon task handling and token efficiency, and updates on August 1, August 27 and September 9 improved general capability, vision understanding, trustworthy research and multimodal understanding.