Wait Which Model?
← Back to directory
ByteDance (Doubao/Seed)

Seed 2.0 Pro

Released Feb 14, 2026 · knowledge cutoff unpublished

Status
Superseded
Location
China
Modality
Multimodal
Context window
256K
Max output
128K
Speed
—
Price ($/MTok in / out)
$0.47 / $2.37
Cost per task
—
Open weights
No
Benchmarks6 of 8 reported

MMLU-Pro

87%

GPQA Diamond

88.9%

SWE-bench Verified

76.5%

AIME (latest)

98.3%

Humanity's Last Exam

32.4%

ARC-AGI-2

37.5%

Terminal-Bench 2.1

—

LMArena Elo

—
Strengths & weaknesses

Strengths

  • Ranked in the top handful of models on human-preference arena testing at launch, reflecting strong everyday conversational quality rather than just benchmark strength
  • Handles long, multi-part Chinese-language instructions — tone control, formatting, few-shot pattern-following — noticeably better than its Seed1.8 predecessor
  • Native video tool-use (VideoCut) lets it re-inspect footage at a higher frame rate mid-task instead of reasoning over static frames alone

Weaknesses

  • ByteDance's own report is unusually candid about real gaps against Claude on repository-scale coding and against Gemini on long-tail factual knowledge
  • Long-context retrieval (needle-style) tasks lag its overall reasoning strength — it can lose track of scattered details in very long inputs
  • Hallucination-robustness trails GPT-5.2 on several factuality checks despite leading on many other evaluations

Officially launched 2026-02-14 (Seed2.0 Model Card, ByteDance Seed, hosted at lf3-static.bytednsdoc.com; announcement blog at seed.bytedance.com). MMLU-Pro 87.0, GPQA Diamond 88.9, AIME 2025 98.3 and HLE 32.4% (no-tools, text-only) are from the model card's Table 3; SWE-bench Verified 76.5% is from Table 11 (agentic capacity evaluation); ARC-AGI-2 37.5 is from Table 3's General Reasoning section. Terminal-Bench figure in the same table (55.8) is version 2.0, not the tracked 2.1 metric, so left null here. LMArena listed only a rank (6th on the Text Arena leaderboard as of 2026-02-16), not an Elo score, so lmarenaElo stays null. Context window (256K) and max output (128K) are corroborated across third-party trackers (CloudPrice, others) citing the API model id 'doubao-seed-2-0-pro-260215'; not independently confirmed on an official ByteDance spec page, so treat as secondary-sourced. Pricing from the model card's Table 1 ($0.47/$2.37 per MTok, quoted alongside ¥3.41/¥17.04).

For developers

API model strings

  • volcenginedoubao-seed-2.0-pro

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor

News