Local Ai
ACE-Step 1.5 XL Base — BF16 version (converted from FP32)
ACE-Step 1.5 XL Base is a 4B-parameter Diffusion Transformer (DiT) decoder in the ACE-Step 1.5 open-source music generation framework, designed to deliver higher audio quality than the earlier 2B-p...
ACE-Step 1.5 XL Base is a 4B-parameter Diffusion Transformer (DiT) decoder in the ACE-Step 1.5 open-source music generation framework, designed to deliver higher audio quality than the earlier 2B-parameter models. The XL weights are approximately 9 GB in BF16 format and require at least 12 GB VRAM with offload enabled, or 20 GB or more without offload. The Reddit post shared by user "local-ai" presents a BF16-precision version of the model converted from the original FP32 weights, halving the file size compared to the ~20 GB FP32 originals, with BF16 likely enabling operation on as little as 8 GB of VRAM.
Related
- Qwen3.5-4B-Base-ZitGen-V1
- LTX 2.3 - Image + Audio + Video ControlNet (IC-LoRA) to Video
- b8753
- ASUS UGen300 USB AI Accelerator 8GB for local inference
Source: local-ai