MXFP4 (Microscaling FP4) is a 4-bit floating point format from the Open Compute Project MX Specification. Unlike integer quantization (Q4_K_M etc.), MXFP4 stores weights as E2M1 floating point values with shared E8M0 block scales.
On NVIDIA Blackwell GPUs (RTX 5090, RTX 5080, RTX 5070 Ti, etc.), MXFP4 models use native FP4 Tensor Core instructions — the hardware decodes and multiplies FP4 values directly, without dequantization overhead. This provides measurable speedups over traditional quantization formats.
On older NVIDIA GPUs (Ampere, Ada Lovelace, Turing), MXFP4 models still work correctly — the engine falls back to INT8 dp4a matmul with software dequantization. You'll get the same quality and memory savings, just without the Tensor Core speed boost.
Benchmark Results
Tested on RTX 5090 (Blackwell) with acestep-v15-merge-base-turbo-xl-ta-0.5:
Quality (Listening Test — Full Quality ✅)
MXFP4 passed subjective listening tests with no audible degradation compared to Q4_K_M and only minimal loss compared to BF16.
Speed (Blackwell Tensor Core Advantage)
Workload
Q4_K_M
NVFP4
IQ4_XS
MXFP4
30s / 60 steps
12.1s
6.0s
6.0s
6.0s
60s / 100 steps
12.1s
14.1s
12.0s
10.0s
120s / 200 steps
22.1s
24.1s
21.1s
18.1s
MXFP4 is 22% faster than Q4_K_M and 33% faster than NVFP4 on compute-heavy workloads — same file size, same bits per weight, purely from native FP4 Tensor Core acceleration.
Memory
Model Size
BF16
Q4_K_M
MXFP4
Compression
Standard
4,570 MB
2,401 MB
1,223 MB
3.7x
XL
9,516 MB
2,851 MB
2,537 MB
3.8x
Available Models
Standard Models (~1.2 GB each)
Filename
Description
acestep-v15-base-MXFP4.gguf
V1.5 base model
acestep-v15-sft-MXFP4.gguf
V1.5 SFT (supervised fine-tuned)
acestep-v15-sftturbo50-MXFP4.gguf
V1.5 SFT + Turbo 50%
acestep-v15-turbo-MXFP4.gguf
V1.5 Turbo (fewer steps needed)
acestep-v15-turbo-continuous-MXFP4.gguf
V1.5 Turbo Continuous
acestep-v15-turbo-shift1-MXFP4.gguf
V1.5 Turbo with shift=1
acestep-v15-turbo-shift3-MXFP4.gguf
V1.5 Turbo with shift=3
XL Models (~2.5 GB each)
Filename
Description
acestep-v15-xl-base-MXFP4.gguf
V1.5 XL base model
acestep-v15-xl-sft-MXFP4.gguf
V1.5 XL SFT
acestep-v15-xl-sftturbo50-MXFP4.gguf
V1.5 XL SFT + Turbo 50%
acestep-v15-xl-turbo-MXFP4.gguf
V1.5 XL Turbo
acestep-v15-merge-base-turbo-xl-ta-0.5-MXFP4.gguf
XL merge: base + turbo (α=0.5)
acestep-v15-merge-sft-turbo-xl-ta-0.3-MXFP4.gguf
XL merge: SFT + turbo (α=0.3)
acestep-v15-merge-sft-turbo-xl-ta-0.7-MXFP4.gguf
XL merge: SFT + turbo (α=0.7)
Usage
With HOT-Step-CPP
Place the .gguf files in your acestep.cpp/models/ directory
Launch HOT-Step-CPP — models appear automatically in the DiT model dropdown
These models are compatible with acestep.cpp and HOT-Step-CPP on any supported GPU. The Blackwell Tensor Core optimisation activates automatically when running on Blackwell hardware — no configuration needed.
GPU Family
Architecture
Works?
Native FP4 TC
Expected Benefit
RTX 5090/5080/5070 Ti/5070
Blackwell
✅
✅ Yes
Fastest — native FP4 Tensor Cores
RTX 4090/4080/4070
Ada Lovelace
✅
❌ Fallback
Same quality + compression, dp4a matmul
RTX 3090/3080/3070
Ampere
✅
❌ Fallback
Same quality + compression, dp4a matmul
RTX 2080/2070
Turing
✅
❌ Fallback
Same quality + compression, dp4a matmul
AMD (ROCm)
RDNA/CDNA
✅
❌ Fallback
Same quality + compression, software path
Tip: On non-Blackwell GPUs, MXFP4 provides the same ~3.8x compression and quality as Q4_K_M, but without the speed advantage. If you're on Ada Lovelace or older and want maximum speed, Q4_K_M is equally good. MXFP4 shines specifically on Blackwell.
This is a native FP4 matrix-multiply-accumulate that operates directly on E2M1 data with E8M0 scales — no dequantization step needed. The result is accumulated in FP32 for numerical stability.
License
These quantized models inherit the license of the original ACE-Step 1.5 project.
Credits
ACE-Step 1.5 — Original model and training by the ACE-Step team
