Skip to content

Next Token다음 조각 Home

Model Architecture & Training Methods Moonshot AIJul 28, 2025

Kimi K2

Kimi K2 — a 1T MoE trained on 15.5T tokens with MuonClip, without a single loss spike

Why it matters

Muon validated at 1T scale

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2507.20534

Added in Week 35, 2026 · See that week

Same week

Aug 2Is Chain-of-Thought Reasoning of LLMs a Mirage?Paper Aug 2AI Act Chapter III Section 4, Chapters V (GPAI), VII and XII start to applyPress Jul 29Elasticsearch 9.1 makes BBQ default for dense vectors of 384+ dims, adds ACORN-1Official

Nearby on the same axis

Jun 17, 2025Inception Labs announces MercuryPaper Sep 19, 2025xAI releases Grok 4 FastOfficial Oct 30, 2025Kimi LinearPaper Dec 4, 2025Google Research introduces Titans and MIRAS, the framework that generalizes itOfficial