ERNIE 4.5 21B A3B Thinking

ERNIE 4.5 21B A3B Thinking has 21.8B parameters. Estimated VRAM is 15.7 GB at Q4, 27.2 GB at Q8 and 50.7 GB at FP16, all estimates from the parameter count, covering weights plus about 15% and 0.5 GB for a short context.

Yes at Q4. A 24 GB GPU and a 16 GB GPU both fit the model at Q4 with an estimated 15.7 GB. A 12 GB or 8 GB GPU does not fit at Q4, Q8 or FP16, estimate.

Yes. Q4 needs an estimated 15.7 GB, and macOS lets the GPU use about 75% of unified memory. The smallest common Mac memory size that fits Q4 is 24 GB, estimate.

Yes. The model has open weights, so you can self-host ERNIE 4.5 21B A3B Thinking for full control of your data.

Run ERNIE 4.5 21B A3B Thinking locally

No. Input modalities are text only, so ERNIE 4.5 21B A3B Thinking does not read images, audio, video or PDFs.

The closest alternatives, matched on model kind and release date with no comparable price, are Apertus 70B by Swiss AI, Apertus 8B by Swiss AI and Qwen3-Next 80B-A3B (Thinking) by Alibaba.

Apertus 70BERNIE 4.5 21B A3B Thinking vs Apertus 70BApertus 8BERNIE 4.5 21B A3B Thinking vs Apertus 8BQwen3-Next 80B-A3B (Thinking)ERNIE 4.5 21B A3B Thinking vs Qwen3-Next 80B-A3B (Thinking)

ERNIE 4.5 21B A3B Thinking has open weights and is licensed under Apache 2.0.

Yes. The license is Apache 2.0, a permissive license that allows commercial use.

The weights are available on Hugging Face in the repository baidu/ERNIE-4.5-21B-A3B-Thinking.

Run ERNIE 4.5 21B A3B Thinking locally

ERNIE 4.5 21B A3B Thinking was made by Baidu and released on September 9, 2025.

More from Baidu

Yes. ERNIE 4.5 21B A3B Thinking has open weights.