Seed2.1 (Doubao Seed 2.1)

Seed2.1 is a model family from ByteDance Seed, announced on the team's blog on 23 June 2026 and made available to users of Doubao and Volcano Engine^[1]. It ships in two versions, Seed2.1 Pro and Seed2.1 Turbo, both positioned for real productivity scenarios rather than demonstrations^[2]. Turbo is identified in the API by the model ID dola-seed-2-1-turbo-260628^[2]. A model card was published alongside the release^[3].

Where it comes from

Seed2.1 builds on Seed2.0. ByteDance describes the new generation as a response to user feedback asking for more reliable answers and more consistent delivery, that is, to the complaint that a model can look strong in a demo yet behave unevenly in day-to-day work^[1]. The gains it lists are in general agent ability, coding delivery, knowledge, reasoning and multimodal understanding^[1][2].

Reported scores

ByteDance gives the following results, Pro first and Turbo second: BeyondAIME 87.0 and 88.0; SuperGPQA 70.8 and 67.4; Terminal Bench 2.1 71.0 and 67.6; MMMU-Pro 81.6 and 80.1^[2].

The same materials place other vendors' models next to these numbers. On BeyondAIME, ByteDance's comparison table lists Claude Opus 4.7 at 79.0 and GPT-5.5 at 91.0; on Terminal Bench 2.1 it lists them at 71.7 and 73.8^[2]. Gemini 3.1 Pro appears at 90.0 on BeyondAIME^[2] and at 76.6 on SuperGPQA, against 70.8 and 67.4 for the two Seed2.1 versions^[2].

Read plainly: Turbo edges out Pro on BeyondAIME (88.0 against 87.0), while Pro leads on the other three benchmarks^[2]. Against the external figures shown, Seed2.1 Pro sits below GPT-5.5 and Gemini 3.1 Pro on BeyondAIME and below Gemini 3.1 Pro on SuperGPQA, while both Seed2.1 versions are above Claude Opus 4.7 and GPT-5.5 on Terminal Bench 2.1 in ByteDance's table^[2].

Agents and video

ByteDance attributes part of the agent improvement to reinforcement learning: in its account, reinforcement learning cut the average number of steps a computer-use agent needed per task by 16%^[1]. In ByteDance's own tests, Seed2.1 Pro scored highest on GDPVal^[1]. On Agents' Last Exam (ALE) the Pro model ranks among the top tier of participating models; Seed2.1 has the highest score on MobileWorld and is competitive on OSWorld^[1].

For video, Pro scores 89.2 on VideoMME, and the models accept inputs including hour-long video^[2].

An earlier appearance

Before the release, a Seed2.1 Preview appeared on the Code Arena frontend leaderboard, ranking 8th with a score of 1539^[1].

How to read the numbers

All the figures above come from ByteDance's own release materials, and the comparison rows are drawn from its table rather than from an independent evaluation. Benchmark scores show where a model stands on a handful of defined tasks; they do not settle whether it fits a particular job.