randmaru/Qwen3.8-27B-mlx-mxfp4-text-generation-only

This is an MXFP4 MLX quantization of Qwen/Qwen3.8-27B for Apple Silicon inference.

Text-generation-only: all 333 vision weights in BF16 were removed.

---

MXFP4 vs 4Bit Quantization Comparison

Parameter	MXFP4	4Bit
Quantization format	4‑bit floating point with microscaling, group 32, shared exponent E8M0	4‑bit integer (INT4/NF4)
Tensor types	U8, U32, BF16	BF16, U32
Parameter size (safetensors)	~14.3 GB	~15.13 GB
Total storage (all files)	~14.32 GB	~15.15 GB
Hardware support	Most efficient on GPUs with microscaling / FP8 tensor core support	Broad support, but often requires specialized INT4 kernels
Apple Silicon compatibility	Designed with hardware microscaling support in Apple Neural Engine / GPU	Works, but without specialized Neural Engine optimization
Inference speed	Higher on compatible hardware: FP path, lower dequantization overhead, higher throughput	Kernel‑dependent; usually lower or comparable at similar quality
Quality	Better preserves dynamic range, less degradation on outliers	Higher risk of accuracy loss on outliers at the same bitrate

Key takeaways:

• Parameter size: The MXFP4 version has a smaller safetensors footprint: ~14.3 GB vs ~15.13 GB for the 4-bit version.
• Total storage: MXFP4 also occupies less total disk space: ~14.32 GB vs ~15.15 GB when summing all repository files.
• Vision weights removed: Both repositories are text-generation-only. All 333 vision weights in BF16 were removed to reduce size and focus inference on text generation.
• Performance: MXFP4 typically delivers higher inference throughput on hardware with microscaling/FP8 support, especially on Apple Silicon. The shared exponent per group of 32 elements reduces dequantization overhead and enables use of floating‑point tensor cores.
• Apple Silicon optimization: MXFP4 is designed with Apple Neural Engine / GPU microscaling support in mind, making it the recommended choice for MacBook.
• Quality: MXFP4 preserves dynamic range better, so generation quality can be higher at the same compression level.

Actual speed depends on the backend, GPU, batch size, and quantization implementation.