---
title: "Metis (FP4 LLM training paper)"
type: "entity"
entity_kind: "concept"
status: "watching"
canonical_name: "Metis"
aliases: ["Metis: Training LLMs with FP4 Quantization"]
first_seen: "2026-08-07T00:00:00.000Z"
writer_model: "claude-sonnet-5"
connects_to: ["anisotropy (singular-value spectra)","FP4 quantization","GIFT","low-bit LLM training"]
seek_code_commit: "649b1a4"
---


"Metis: Training LLMs with FP4 Quantization" (arXiv:2509.00404; Fudan University / University of Bath / Oxford Suzhou Centre / Shanghai Innovation Institute / Huawei; submitted 2025-08-30) diagnoses anisotropic singular-value spectra across weight, activation, and gradient matrices as a barrier to low-bit LLM training, and fixes it with spectral-domain partitioning rather than a Fisher/K-FAC coordinate transform. It entered the vault as the closest independent precedent for [[claim-gift-2026-gradient-anisotropy-isotropic-transform|GIFT]]'s anisotropy diagnosis — related in theme, distinct in mechanism, model, precision target, and metric, and roughly ten months earlier. Watching rather than a full hub: one capture's worth of material so far; its own headline claim (beating Nvidia's unreleased NVFP4 recipe on LLaMA-3 8B) is flagged as a separate future capture, which would be the natural trigger to graduate this to `hub`.

## References
- [[claim-metis-2025-anisotropy-diagnosis-predates-gift-not-a-replication]]