- · claude-code · mcp
Moving houtini-lm to vLLM: What I learned
I decommissioned Hopper (my local LLM bootstrapped server) and moved my local models to a two-card 4090 rig on vLLM. Much faster - but the tool I wrote to save Claude tokens started handing back blank answers, and it looked like the new rig's fault. It wasn't. Here's the honest version, and how v3.2.1 fixes it.
Read article - · local-llm · gpu
Tuning vLLM on a Modded 48GB RTX 4090: From 18 to 133 Tokens per Second
Three days of benchmarking vLLM on a modded 48GB RTX 4090: the quantisation trap that costs Ada owners 40% of their speed, the free 1.9x from speculative decoding, the KV compression that never served once, and what cutting 120 watts costs you. Every number measured, every failure kept.
Read article - · claude-code · mcp
Giving Claude a Local Sidekick: a vLLM Journey Update
I want Claude to have a powerful local sidekick through houtini-lm. I'm not there yet, but this month I got noticeably closer - a vLLM rebuild, a benching spree on the 48GB 4090, and a face-off that deleted two models in an evening.
Read article - · claude-code · openrouter
How I Cut My Claude Code Token Spend with Houtini LM (Architect, Drafter, QA)
Claude as the architect. A cheap local model as the drafter. Claude QAs everything. The token bill drops as a side-effect of using each model for what it does best. Here is how I run it on my own rig.
Read article