Model Details
ERNIE 4.5
Quick answer: ERNIE 4.5 is Baidu's April 2025 model scoring 60.7% ARC-E, 40.7% C-Eval, and 40.6% ARC-C. Proprietary API — weak on standard benchmarks but intended for Chinese-market use cases.
At a Glance
Where ERNIE 4.5 leads
- Chinese-language optimization (C-Eval 40.7%)
- Baidu ecosystem integration
- ERNIE generation flagship at launch
Where it lags
- 60.7% ARC-E, 40.6% ARC-C — below GPT-3.5 level on English reasoning
- 25.2% GSM8K — poor math performance
- Proprietary: no open weights
Best for: Baidu ecosystem / Chinese-market deployments; use only when required by platform integration.
What ERNIE 4.5 Is
ERNIE 4.5 is Baidu's April 2025 Chinese large language model. The benchmark scores are unusually low for a 2025 flagship — suggesting ERNIE's primary optimization target is Chinese-language tasks and Baidu product integration, not standard English benchmarks.
Specifications
| Field | Value |
|---|---|
| Organization | Baidu |
| License | Proprietary (API only) |
| Release date | April 2025 |
| Modality | Text only |
Pricing
Available via Baidu Qianfan API.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| ARC-E | 60.7% | Benchgen evaluation | 2025-04 |
| C-Eval | 40.7% | Benchgen evaluation | 2025-04 |
| ARC-C | 40.6% | Benchgen evaluation | 2025-04 |
| GSM8K | 25.2% | Benchgen evaluation | 2025-04 |
ERNIE 4.5 vs Alternatives
| Model | ARC-C | C-Eval | License |
|---|---|---|---|
| ERNIE 4.5 | 40.6% | 40.7% | Proprietary |
| Qwen2.5 72B Instruct | — | — | Apache 2.0 |
| GLM-5 | — | — | Proprietary |
For Chinese-language tasks: Qwen2.5 or GLM series outperform ERNIE on standard benchmarks while also serving Chinese-language use cases.
Related Models & Benchmarks
- Compare: Qwen2.5 72B Instruct · GLM-5
- Benchmarks: ARC-C · GSM8K
Frequently Asked Questions
What is ERNIE 4.5?
Baidu's April 2025 model scoring 60.7% ARC-E, 40.7% C-Eval. Proprietary API — Chinese-market focus with low English benchmark scores.Specs from Baidu's ERNIE 4.5 release (April 2025) and Benchgen evaluations. Last updated 2026-07-24.