profile

CctoctoFX

Yuxiao Huang

Software engineer working at the intersection of AI infrastructure and ML systems.


System architecture Large language models AI infrastructure Technology & geopolitics

Grounded in systems thinking — no fluff, just signal.

MiniMax-M3 架构 QA

基于 MiniMax-M3 主报告的配套 QA。覆盖 MSA 多阶段稀疏注意力、GQA 配置、MoE 路由、视觉编码器、MTP 投机解码、训练体系等核心主题。

June 15, 2026 · 37 min · 7795 words · Me

MiniMax-M3 架构深度拆解

MiniMax-M3 是 MiniMax 的旗舰多模态 MoE 模型。核心创新为 MSA(Multi-stage Sparse Attention)索引机制实现长上下文稀疏注意力、GQA(64Q/8KV)、256E MoE(k=8, sigmoid+bias)、视觉编码器(ViT+MLP投影)、MTP×1 投机解码。本期完整拆解 MSA 三阶段索引、MoE 路由数据流、视觉编码架构及训练体系。

June 15, 2026 · 24 min · 4928 words · Me

Kimi-K2.5 架构 QA

基于 Kimi-K2.5 主报告的配套 QA。覆盖 MLA 潜注意力压缩、MuonClip 优化器、MoE 路由等核心主题。

June 12, 2026 · 12 min · 2521 words · Me

Kimi-K2.5 架构深度拆解

Kimi-K2.5 是 Moonshot AI 的旗舰 MoE 模型。核心创新为 MLA 潜注意力压缩、MuonClip 优化器(Muon + 梯度裁剪)、大规模 MoE 路由等。本期完整拆解整体架构、MLA 机制、MuonClip 设计及与同期模型的对比。

June 12, 2026 · 17 min · 3507 words · Me

Nemotron-3-Ultra 架构 QA

基于 Nemotron-3-Ultra 主报告的配套 QA。覆盖 Mamba-2 SSD 混合架构、LatentMoE 潜空间路由、MTP 投机解码、训练体系等核心主题。

June 12, 2026 · 33 min · 6950 words · Me