Fitting a 27B Model Into 16 GB, and the Four Tuning Ideas That Lost
A 12.1 GB uncensored coder model on an RTX 5070 Ti, 40,960 tokens of context, and four tuning attempts that measured slower than the defaults.
Hi, I’m mistercyber.
I build things with AI, Linux, code, and whatever catches my curiosity. This is where I write them down.
Latest
Writing.
Starlink UK Review: I Didn't Think It Would Work at My House — Four Months Later, I'm Still Using It
A sceptic's four-month Starlink case study from a Norfolk house, after a five-year EE G.Fast line was set to become a ~64 Mbps service. One household's unverified experience of a 2.5% obstructed sky, gaming, downloads and an AI workflow, with missing measurements left visible.
I Turned My CachyOS PC Into a Practical AI Coding Workstation
How I combined CachyOS, OpenCode, llama.cpp, a local 9B model, OpenRouter, MCP tools and strict Git checkpoints without turning my gaming PC into an AI-only server.