The Tensormux journal

Notes from the builders.

Experiments, benchmarks, and ideas from the people building inference infrastructure.

Benchmark2026-07-03

Zero failures. Half the latency budget. $0.32 a million tokens.

Our first H100 benchmark, fully reproducible. Tensormux serving Llama-3.1-8B-Instruct on 4×H100: how the control plane held its SLA under sustained production load.

By Nur & KrishRead article

Article2026-06-27

Everything you need to know about Speculative Decoding Inference

A deep dive into speculative decoding — how draft models, EAGLE, Medusa, and lookahead decoding speed up LLM inference without changing the model itself.

By mohitRead article