ERNIE 4.5 VL 424B A47B PT
baidu · released 2025-06-30 · apache-2.0 license
ERNIE-4.5-VL-424B-A47B is Baidu's flagship open-weight vision-language MoE (424B total, 47B active, June 2025, Apache-2.0). Baidu publishes its VL benchmarks only as chart images, so no per-metric numbers could be sourced.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 423.53B total |
| Architecture | ernie4_5_moe_vl |
| Context window | 131K tokens |
| Knowledge cutoff | 2025-03-31 |
| Modalities | text, vision |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | — |
|---|---|
| SWE-bench Verified | — |
| AIME | — |
| MMLU-Pro | — |
| BFCL v3 (tool use) | — |
| Composite score | — |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 247.1 GB | 245.6 GB | — | 131K |
Strengths & weaknesses
Strengths: Large multimodal MoE (424B total / 47B active) with strong visual, document and chart understanding; Thinking and non-thinking modes, 131K context; Apache-2.0
Weaknesses: Vision-language focus; published text benchmark coverage is thin; Very large, heavy to deploy