Seed2.0 (Doubao-Seed-2.0)

Seed2.0, also called Doubao-Seed-2.0, is a model series that ByteDance released on 14 February 2026. It was made available in the Doubao app and through Volcano Engine's API ^[1][2]. The series launched with three general-purpose agent models: Seed2.0 Pro, Seed2.0 Lite and Seed2.0 Mini. ByteDance positioned Pro for long-chain reasoning and agent tasks, and compared it with GPT-5.2 and Gemini 3 Pro ^[3][1]. The company also claimed that token prices were about an order of magnitude below those of comparable models ^[1]. This framing put the family in the agent-model category: models intended to handle multi-step tasks rather than only single-turn questions.

Reported results for Pro

ByteDance reported a set of benchmark scores for Seed2.0 Pro (0215). On GPQA Diamond it scored 88.9; on BeyondAIME, 86.5; on Humanity's Last Exam without tools, 32.4; and on BrowseComp, 77.3 ^[3]. For coding, the reported scores were 46.9 on SWE-bench Pro, 71.7 on SWE Multilingual and 55.8 on Terminal Bench 2.0. Multimodal results included 78.2 on MMMU-Pro and 89.5 on VideoMME ^[3]. On GDPval, a benchmark of real-world professional tasks, Pro scored 54.4 ^[3]. These figures cover several different evaluation areas, from reasoning and coding to multimodal understanding and professional work.

Lite and Mini

Seed2.0 Mini was optimized for inference throughput and deployment density, aimed at high-concurrency and batch generation ^[3]. Within the same family, its emphasis is therefore on handling many requests or large batches rather than on the largest single-response capability.

Seed2.0 Lite received an upgraded version, Seed2.0 Lite (0428), at the end of April 2026. It became the first omni-modal understanding model in the Seed series, taking video, image, audio and text. It scored 88.4 on GPQA Diamond and 64.4 on OSWorld-Verified ^[3]. Its coding scores were 66.6% on SWE Multilingual, 46.6% on SWE-Bench Pro and 43.3% on Terminal Bench 2.0 ^[3]. On OSWorld-Verified, the comparison table on ByteDance's page placed it behind Claude Opus 4.7 at 78.0%, Claude Sonnet 4.6 at 72.5% and GPT-5.4 High at 75.0% ^[3].

Place in the Seed series

Seed2.0 arrived two days after the Seedance 2.0 video model debuted on 12 February 2026 ^[2]. The series was succeeded by Seed2.1 in June 2026 ^[4]. That places Seed2.0 in a rapid sequence of releases: a model family announced close to the video model, then succeeded within months. Its public profile rested on agent tasks, a range of model sizes, reported benchmark results and lower token pricing.