MXFP4
A 4-bit floating-point format that shares one scaling factor across a small block of weights, keeping more dynamic range than plain 4-bit integers. It is what GPT-OSS-120B and Kimi K3 ship in. The floating-point layout means the 16 available values are spread more usefully than evenly spaced integers.
MXFP4 belongs to the number-format family, next to BF16 and FP16, not to the file-format family with GGUF and MLX. A GGUF file can hold MXFP4 weights; the two answer different questions.
What makes it worth a name of its own is that newer GPUs compute it natively rather than unpacking it to something wider first, and that it turns up in weights as released rather than only in community squeezes.
Kimi K3 is the clearest case of the second part. Moonshot trained it quantization-aware in MXFP4 through the whole post-training stage, supervised fine-tuning and reinforcement learning both, so the model was adapting to the 4-bit numerics while it learned rather than being squeezed into them afterwards. Only the expert weights are MXFP4. Activations are MXFP8 and the non-expert layers stay wider. The result ships at 1.56 TB where the same model in BF16 would be about 5.6 TB.