ERNIE 4.5 21B A3B Thinking
Baidu · 2025-09-09 · 21.8B parameters
ERNIE 4.5 21B A3B Thinking is Baidu's open-weight reasoning checkpoint of the ERNIE 4.5 21B A3B text MoE, released in September 2025 after a further round of post-training aimed at thinking quality and depth. It keeps the base architecture (28 layers, 20 query and 4 KV heads, six of 64 routed experts plus two shared experts, about 3B activated parameters per token) and adds a longer thinking budget, tool use and 128K long-context understanding. Weights are published in Transformers format under Apache 2.0, and the model always reasons before answering.
Benchmark scores
| Benchmark | Score |
|---|---|
| Eval Connections | 0.0 |
| Eval Connections — Guess Accuracy | 18.2 |
| SnakeBench Average Apples | 1.9 |
| SnakeBench Best Apples | 5.0 |
| SnakeBench Rating | 18.4 |
| SnakeBench Total Apples | 30.0 |
| SnakeBench Win Rate | 26.7 |