Back

The Frontier Lab

Ghost
thefrontierlab.ai· weekly· Published 2 months ago

A hands-on engineering newsletter benchmarking and troubleshooting large language models run locally on a Strix Halo mini PC with llama.cpp. It reports measured results, including corrections of the author's own earlier advice, on VRAM allocation, speculative decoding and KV cache quantization.

llama.cppstrix halolocal llmspeculative decodingkv cache quantizationvram carveoutrocmvulkanmixture of expertsself-hosted aibenchmarks
Platform
Ghost
Frequency
Weekly
Latest issue
Aug 27, 2026
Language
English
Written for
Engineers and enthusiasts running large LLMs on local hardware
Tone
analytical
Tracked since
Oct 23, 2026

Website: https://thefrontierlab.ai/