ERNIE-4.5-21B-A3B
ERNIE-4.5-21B-A3B is an open-weights language model from Baidu, released June 29, 2025. It has 21B parameters (mixture-of-experts, 3B active).
In this report, we introduce ERNIE 4.5, a new family of large-scale multimodal models comprising 10 distinct variants. The model family consist of Mixture-of-Experts (MoE) models with 47B and 3B active parameters, with the largest model having 424B total parameters, as well as a 0.3B dense model.
Check it against your own PC →
Can you run ERNIE-4.5-21B-A3B locally?
The smallest common GPU it runs on is 6 GB of VRAM (with 32 GB of RAM): Q6_K, GPU + experts in RAM, roughly 18 tokens per second. Estimates assume 32 GB of DDR5 RAM, a 16K context and llama.cpp-style GGUF quantization. Enter your exact hardware in the live app.
| GPU memory | Verdict | Best quant | VRAM | RAM | tok/s |
|---|---|---|---|---|---|
| 6 GB VRAM | Runs · GPU + CPU experts | Q6_K (≈ lossless) | 5.2 GB | 15.2 GB | ~18 |
| 8 GB VRAM | Runs · GPU + CPU experts | Q6_K (≈ lossless) | 7.2 GB | 13.2 GB | ~21 |
| 12 GB VRAM | Runs · GPU + CPU experts | Q6_K (≈ lossless) | 11.2 GB | 9.2 GB | ~27 |
| 16 GB VRAM | Runs on GPU | IQ4_XS (good) | 14.3 GB | – | ~160 |
| 24 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 20.4 GB | – | ~190 |
| 32 GB VRAM | Runs on GPU | Q6_K (≈ lossless) | 20.4 GB | – | ~300 |
Benchmark results
Epoch AI has not published results for ERNIE-4.5-21B-A3B yet. New models are usually evaluated within days to weeks.
Links
More from Baidu
ERNIE 5.5 Flash Base · ERNIE 5.5 Flash Base BF16 · ERNIE 5.1 · ERNIE-4.5-VL-28B-A3B · ERNIE-4.5-300B-A47B · ERNIE-4.5-0.3B