OpenAI released its open-weight gpt-oss Mixture of Experts (MoE) models with expert weights natively quantized in MXFP4, alongside day-zero support in llama.cpp and GGUF. Shipping native MXFP4 weights let a 120B-parameter MoE model fit inside a single 80 GB GPU without post-hoc quantization loss.
FP4
Also known as: 4-bit Floating Point · E2M1 · MXFP4 · OCP MXFP4
Plain English
A 4-bit floating-point number format that uses uneven spacing—finer near zero, coarser for large values—making 4-bit AI weights and activations more accurate than uniform 4-bit integers and executable directly on modern AI chips.
Technical Overview
FP4 typically uses the E2M1 layout (1 sign bit, 2 exponent bits, 1 mantissa bit), representing 16 non-uniform values that better match bell-curved neural weight distributions than INT4. Because 4 bits alone lack dynamic range, practical deployments rely on block-scaled formats such as the Open Compute Project's MXFP4 (32-element blocks sharing an E8M0 power-of-two scale) or NVIDIA's NVFP4.
Chronological Entries (3)
NVIDIA announced its Blackwell GPU architecture with a second-generation Transformer Engine adding native hardware FP4 Tensor Core execution. Native 4-bit floating-point matrix math doubled peak compute throughput over FP8 and laid the hardware foundation for both MXFP4 and NVFP4 serving.
Researchers from the OCP Microscaling Alliance published Microscaling Data Formats for Deep Learning, evaluating the MXFP4 standard (E2M1 elements paired with a shared 8-bit E8M0 scale per 32-value block). Sharing one exponent scale across 32 elements let 4-bit floating-point dot products execute in hardware with only 0.25 bits of scale overhead per value.
Direct Relations (4)
- Part ofQuantization
- Alternative toINT4
- Alternative toFP8
- EnablesInference Engines