Marvin G6 | Engineering, AI & Automation

Sign in Subscribe

Topic

LLM inference

A collection of 2 issues

The Frontier Model Monoculture Is Dead: How to Build Inference Pipelines That Adapt

Cut LLM inference costs 50-70% by routing tasks to the right model tier, cascading on quality signals, and keeping model names out of application code.

The Hidden Infrastructure Tax: Running GLM-5.2, Kimi K2.7, and Other Frontier Open-Weight Models in Production

Self-hosting GLM-5.2 or Kimi K2.7? Learn the real infrastructure costs: KV cache sizing, GPU networking, serving frameworks, and a framework for deciding when to self-host.