ByteDance releases Seed2.1 agent-focused models
ByteDance's Seed team has released Seed2.1, a family of agent-capable models aimed at real-world productivity, and says its Pro variant has the highest score on the GDPVal work-task benchmark.
ByteDance's Seed team announced the Seed2.1 model family on June 23, 2026. The models are available as Doubao Seed 2.1 to Doubao and Volcano Engine users.
The team describes Seed2.1 as a response to user feedback after Seed2.0, which asked for more reliable answers and more consistent results. ByteDance lists three areas of improvement: general agent work across tools and environments, end-to-end coding, and multimodal understanding, including complex visual inputs and video. For coding, the company names requirement analysis, feature implementation, bug fixing, environment setup and result validation in enterprise development.
On benchmarks, ByteDance says Seed2.1 Pro has the highest score on GDPVal, which measures how well models complete real-world work tasks and the economic value of that work. The company also says Seed2.1 Pro ranks "among the top tier" on Agents' Last Exam (ALE). Because ALE is recent, ByteDance argues it is a fairer test of how well models generalize. The post describes results on Workspace Bench, Agent Startup Bench, xDailyBench and Doubao Multi-Turn Bench as "consistent" or "steady," and says the model "remains competitive" on Toolathlon and ClawBench. Several of these benchmarks, including SeedClawBench and Doubao Multi-Turn Bench, are internal to ByteDance.
The excerpt available does not give exact scores, model sizes, pricing or API details. Apart from the Pro tier, it also does not say which variants make up the family.