Definition
Microscaling FP4 weight format used in quantization-aware training and inference for large MoE models such as kimi-k3 (paired with MXFP8 activations).
Key Points
- moonshot-ai trains kimi-k3 with MXFP4 weights from SFT onward (2026-07-26-kimi-k3-moonshot-official-blog)
- vllm integrates SiTU-enabled MXFP4 MoE paths for day-0 K3 serving (2026-07-26-vllm-kimi-k3-preview)
- Secondary sources conflict on download size (~1.4TB class vs ~594GB prior K2 packs) — verify model card at drop
Warning
Prefer official model card for storage/VRAM figures after July 27, 2026 weight release.