ERNIE 4.5 21B A3B
ERNIE 4.5 21B A3B is Baidu's language model with a 120K context window and up to 8K output tokens, starting at $0.07 / 1M input and $0.28 / 1M output. A 21B-parameter Mixture-of-Experts LLM from Baidu with 3B active parameters per token, featuring multimodal heterogeneous architecture for cross-modal knowledge fusion.
Capabilities
Input1/5
Text✓
Image·
Audio·
Video·
PDF·
Output1/5
Text✓
Image·
Audio·
Video·
Embedding·
Capabilities2/13
Reasoning·
Adaptive Reasoning·
Function Calling✓
Parallel Function Calling✓
Structured Outputs·
Native JSON Schema·
Web Search·
URL Context·
Computer Use·
Code Execution·
File Search·
Prompt Caching·
Assistant Prefill·
Pricing by Provider
US Dollar ($)
Per 1M tokens
| Provider | Standard | |
|---|---|---|
| Input $ / 1M | Output $ / 1M | |
| $0.07 | $0.28 | |
Cost Calculator
US Dollar ($)
Preset:
Cheapest Instances to Run It
Cloud GPU instances that can host ERNIE 4.5 21B A3B, ranked by cheapest on-demand price. The model needs about 50 GB of GPU memory at FP16 precision (estimated from its parameter count), so treat the fit as guidance rather than a guarantee.
All clouds
FP16 (full precision)
US Dollar ($)
Instance | Cloud | GPU | VRAM | Price | Cheapest region | |
|---|---|---|---|---|---|---|
| Standard_NP10s |  Azure | AMD Alveo U250 FPGA (64GB) | 64 GB | $1.65/hr | westus2 | |
| Standard_NV32as_v4 | Azure | AMD Radeon Instinct MI25 | 64 GB | $1.86/hr | westus2 | |
| Standard_NP20s | Azure | 2× AMD Alveo U250 FPGA (64GB) | 128 GB | $3.30/hr | westus2 | |
Versions
| Version | Released | Context | Input / 1M | Output / 1M | Status |
|---|---|---|---|---|---|
| ERNIE 4.5 21B A3B | 120K | $0.070 | $0.280 | Current | |
| ERNIE 4.5 21B A3B PT | — | 4K | $0.100 | $0.100 | Available |
| ERNIE 4.5 300B A47B PT | — | 4K | $0.100 | $0.100 | Available |