ERNIE 4.5 VL 28B A3B Thinking

At Q4 it needs an estimated 21.2 GB of VRAM, at Q8 about 36.7 GB and at FP16 about 68.7 GB. All figures are estimates from the 29.7B parameter count, covering weights plus about 15% and 0.5 GB for a short context.

Yes on a 24 GB GPU, it fits at Q4 with an estimated 21.2 GB. On a 16 GB, 12 GB or 8 GB GPU it does not fit at Q4, Q8 or FP16, per the estimates.

Yes. At Q4 it needs an estimated 21.2 GB, and assuming macOS lets the GPU use about 75% of unified memory, the smallest common Mac memory size that fits is 32 GB.

Yes, the model has open weights, so you can self-host it for full control of your data.

Run ERNIE 4.5 VL 28B A3B Thinking locally

It supports text and image inputs, but not audio, video or PDFs.

Kimi K2 Thinking Turbo by Moonshot AI, DR Tulu 8B by Ai2 and Olmo 3 32B Think by Ai2. They are matched on model kind and release date, with no comparable price.

Kimi K2 Thinking TurboERNIE 4.5 VL 28B A3B Thinking vs Kimi K2 Thinking TurboDR Tulu 8BERNIE 4.5 VL 28B A3B Thinking vs DR Tulu 8BOlmo 3 32B ThinkERNIE 4.5 VL 28B A3B Thinking vs Olmo 3 32B Think

It is open weights, released under the Apache 2.0 license.

Yes, the Apache 2.0 license is permissive and allows commercial use.

The weights can be downloaded from the Hugging Face repository baidu/ERNIE-4.5-VL-28B-A3B-Thinking.

Run ERNIE 4.5 VL 28B A3B Thinking locally

Baidu made it, and it was released on November 11, 2025.

More from Baidu

Yes, it is free to use as open weights.