ByteDance Unveils Doubao 2.0 Pro: A Domestic Model Taking on GPT‑5.2

ByteDance announced the Seed 2.0 series, with the flagship Seed 2.0 Pro (also called Doubao 2.0) positioned as a “special‑forces” model aimed at solving real‑world complex problems and competing with GPT‑5.2, Claude Opus 4.5 and Gemini 3 Pro.

The model family includes Pro, Lite and Mini variants plus a dedicated Code model; Seed 2.0 Pro represents the highest capability.

1. Math and Vision: Breaking the “subject‑bias” Curse

On math‑reasoning benchmarks such as MathVista and MathVision, Seed 2.0 Pro achieves industry‑leading scores, directly matching Gemini 3 Pro High. In visual perception tests like VLMsAreBiased, it obtains the top score, indicating accurate visual understanding without severe hallucinations.

2. Document Understanding

Seed 2.0 is optimized for unstructured information. On ChartQAPro and OmniDocBench 1.5 it reaches top‑tier performance, demonstrating not just OCR but genuine logical comprehension of PDFs, financial tables and scanned documents.

3. Long Video and Long Text: Remarkable Memory

Evaluated on VideoMME and LongVideoBench, Seed 2.0 Pro excels, even surpassing human performance on several metrics, making it suitable for video analysis and meeting‑summary generation.

4. Science and Code: Pushing Human‑Level Reasoning

On STEM benchmarks such as FrontierSci, Seed 2.0 Pro scores explosively, beating Gemini 3 Pro in some scenarios and earning gold‑medal results on ICPC‑style programming and IMO‑style math tests. It can also draft biological experiment plans, moving toward an “AI scientist”.

5. Complex Instruction Following

In Instruction‑Following evaluations, Seed 2.0 shows strong consistency, a key factor for agents to stay on track during long‑chain tasks.

6. Deep Research and Real‑World Tasks

On benchmarks like GDPVal‑Diamond and XPert Bench, Seed 2.0 Pro ranks among the industry’s top tier, handling customer‑service QA, information extraction and intricate B‑to‑B workflows.

With the model now available in the Doubao app (Expert mode) and the IDE tool Trae (Doubao‑Seed‑2.0‑Code), users can start using it immediately. Token inference cost has dropped by roughly an order of magnitude, making the technology more affordable and accessible.

Overall, ByteDance’s release showcases not only raw performance but a deep understanding of agent‑centric AI: models must be smart, reliable, and capable of tackling genuine complex problems.