Hands-on: Doubao-Seed-Evolving touts 1M context and leaner tool use
A Juejin author published a hands-on test of Doubao-Seed-Evolving, a new Doubao Seed model that the post says extends the context window to 1M tokens, holds up better on longer multi-step tasks with complex dependencies, and uses fewer tokens and tool-call rounds than Doubao-Seed-2.1-pro in internal Coding Agent and LLM/VLM evaluations. The author wired the model into Claude Code through the Volcano Ark-compatible endpoint and used a single prompt to build an accounting skill with horoscope-based emotional feedback, then compared it with a Seed-2.1-pro version of the same task and found the Evolving build offered fuller page-based interaction. The post also notes that ledger and personal data can be stored in a local knowledge base for long-term records.
It is one of the few early hands-on accounts of Doubao-Seed-Evolving, though the efficiency gains cited come from vendor internal benchmarks and the author's comparison is a single subjective test without quantitative metrics.
1. Doubao-Seed-Evolving来袭！
豆包 Seed 这次新上的 Doubao-Seed-Evolving不错，主要有以下优点：
第一个是支持 1M 超长上下文。 模型上下文窗口扩展到了 1M tokens，这对我来说意味着我可以用它写长篇小说、剧本了（一直在做漫剧从未放弃）。
第二个是长程任务能力提升。 它在持续时间更长、步骤更多、依赖关系更复杂的任务里表现会更稳定一些，这意味着可以用它来开发更加复杂的系统。
第三个是 Tokens 效率提升。 在内部 Coding Agent 评测，以及 LLM、VLM 基础能力评测里，对比 Doubao-Seed-2.1-pro，Seed Evolving 的 tokens 消耗更少，工具调用轮次也更简洁。
#Doubao-Seed-Evolving#Doubao#ByteDance#Doubao-Seed-2.1-pro#Claude Code#Volcano Ark#Obsidian