Marcos Hernanz

Software engineer working on language-model training. Previously on the Next.js team at Vercel, building coding-agent and evaluation infrastructure.

Email · CV · GitHub · LinkedIn · X

Research

• How the Optimizer Shapes Routing: Load Balancing, Specialization, and Stability in Mixture-of-Experts Language Models Trained with Muon (bachelor's thesis, in progress)
• Multi-Head mHC: No Gain at Equal Parameters (preprint, 2026) · pdf
• The Optimal Rate Constant of Heavy-Ball Momentum on Smooth Strongly Convex Functions (preprint, 2026) · pdf

Writing

• How we closed 1,500 GitHub issues in one month (Next.js blog, 2026)

Projects

• llm-internals — DeepSeek-V4.1-Flash, Triton kernels and post-training, from scratch
• llm-lab — a BPE tokenizer, a 39.85B-token TPU run, and one modern mechanism per run
• BareTensor — a PyTorch-style tensor and autograd library in C++