Wait Which Model?
← Back to directory
Baidu (ERNIE)

ERNIE 5.0

Released Jan 22, 2026 · knowledge cutoff unpublished

Status
Superseded
Location
China
Modality
Multimodal
Context window
128K
Max output
66K
Speed
—
Price ($/MTok in / out)
$0.60 / $2.10
Cost per task
—
Open weights
No
Benchmarks5 of 8 reported

MMLU-Pro

83.8%

GPQA Diamond

86.36%

AIME (latest)

89.06%

Humanity's Last Exam

25.81%

LMArena Elo

1447

SWE-bench Verified

—

Terminal-Bench 2.1

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • First Chinese model to break the LMArena Text global top 10 at launch, on par with GPT-5.1 High and ahead of Gemini 2.5 Pro at the time
  • Native full-modal (text/image/audio/video) unified modeling rather than a bolted-on vision encoder — leads on document- and chart-understanding benchmarks like OCRBench, DocVQA and ChartQA
  • Trillion-parameter MoE activates under 3% of parameters per inference, keeping serving cost down despite the 2.4T total size

Weaknesses

  • Real-world testing found it ignoring explicit format instructions (e.g. firing the image generator when told to output SVG code only), pointing to instruction-following gaps behind its benchmark scores
  • Developers reported tool-call loops and rough edges running it in agentic workflows despite strong raw capability
  • Baidu discloses no knowledge-cutoff date

Officially launched 2026-01-22 at the Ernie Moment Conference 2026 Shanghai (a preview build, ERNIE-5.0-Preview, had circulated from November 2025). Benchmark figures (MMLU-Pro 83.80, GPQA-Diamond 86.36, AIME 2025 89.06, HLE 25.81) are Baidu's own post-trained-model numbers from the ERNIE 5.0 Technical Report (arxiv.org/abs/2602.04705, Table in the post-training results section) — self-reported and not independently reproduced; the same report separately lists lower base-model-only scores (MMLU-Pro 75.58, GPQA-Diamond 57.30) which are not used here. lmarenaElo 1447 is the official arena.ai Text leaderboard score for the ERNIE-5.0-0110 checkpoint (independent measurement, preferred over Baidu's own '#8 globally / 1460' launch-day figure, which the leaderboard has since revised). SWE-bench Verified and ARC-AGI-2 were not reported by Baidu or found on any independent leaderboard. Pricing ($0.60/$2.10 per MTok) and context/output figures (128K/65,536) are corroborated across third-party trackers citing Baidu Qianfan's console, not confirmed on an official Baidu pricing page directly. Closed weights; no license record.

For developers

API model strings

  • baidu-qianfanernie-5.0

Licence

Proprietary — weights not released

Retirement

No retirement announced

Lineage

No recorded predecessor

News