// Baidu ERNIE · Text LLM
ERNIE-4.5-21B-A3B-PT: VRAM & local deployment
21.9B total · 3B active (MoE) parameters. About 15.6 GB at 4-bit and 49.3 GB at BF16 with one 8K-token request.
released 2025-06-28licence: Commercial OKmemory figures are estimates
01Specs
| Publisher | Baidu · baidu |
|---|---|
| Family / series | Baidu ERNIE · ERNIE-4.5 |
| Type | Text LLM (text-generation) |
| Parameters | 21.9B total · 3B active (MoE) Source: counted from official safetensors weights |
| Context length | 128K tokens (131,072)max_position_embeddings in config.json |
| Architecture | 28 layers · GQA · MoE |
| Released weights | BF16 · 43.9 GBdownload |
| Licence | Apache-2.0 · Commercial OKPermissive licence (Apache-2.0 / MIT / CC-BY / OpenMDW): commercial use allowed. |
| Released | 2025-06-28 (Hugging Face repo created) |
02How much VRAM it needs (estimate)
Total = weights + 10% + 0.5GB overhead + KV cache for one request (FP16 KV). More concurrent users multiply the KV column.
| Precision | Weights | Total @ 8K | Total @ 32K | Total @ 128K | Smallest GPU | Mac |
|---|---|---|---|---|---|---|
| BF16 / FP16 (original) | 43.9 GB | 49.3 GB | 50.7 GB | 56.3 GB | 2 × RTX 5090 (64GB) | Mac 128GB |
| FP8 / INT8 / Q8_0 | 23.3 GB | 26.6 GB | 28.0 GB | 33.7 GB | RTX 5090 32GB | Mac 64GB |
| 4-bit (Q4_K_M / AWQ / GPTQ-Int4) | 13.3 GB | 15.6 GB | 17.0 GB | 22.7 GB | RTX 4090 24GB | Mac 36GB |
KV cache per request: 8K 0.5 GB · 32K 1.9 GB · 128K 7.5 GB
Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →
03How to run it
Only runtimes the publisher or the runtime’s official library names for this model.
- vLLM — mentioned in the official model card
- Transformers — mentioned in the official model card
04Official repositories & quantised builds
1 official repo(s) from baidu.
- ERNIE-4.5-21B-A3B-PT Base
05Other Baidu ERNIE models
All 11 →| Model | Params | 4-bit total | Released |
|---|---|---|---|
| ERNIE-4.5-0.3B-PT · same series | 361M | 0.9 GB | 2025-06-28 |
| ERNIE-4.5-300B-A47B-PT · same series | 300.5B / 47B | 203 GB | 2025-06-28 |
| ERNIE-Image | 8.03B | 5.9 GB | 2026-04-07 |
| ERNIE-Image-Turbo | 8.03B | 5.9 GB | 2026-04-02 |
| ERNIE-4.5-VL-28B-A3B-Thinking | 29.7B / 3B | 20.8 GB | 2025-11-07 |
| ERNIE-4.5-21B-A3B-Thinking | 21.8B / 3B | 15.5 GB | 2025-09-08 |
| ERNIE-4.5-300B-A47B-Paddle | 300.5B / 47B | 203 GB | 2025-06-28 |
| ERNIE-4.5-VL-28B-A3B-PT | 29.4B / 3B | 20.6 GB | 2025-06-28 |
| ERNIE-4.5-VL-424B-A47B-PT | 423.5B / 47B | 285 GB | 2025-06-28 |
| ERNIE-4.5-VL-424B-A47B-Paddle | 423.5B / 47B | 285 GB | 2025-06-28 |
06Sources
Data checked 2025-11-26. Parameters, context and licence are copied from the official repository; memory figures are our estimates.