Writing
Articles
Rounds report what a batch of tests found. Education explains the vocabulary. Articles are the longer arguments that cut across both — the things this bench learned the hard way and would rather you did not have to.
LM Studio Silently Drops Reasoning Effort on Sideloaded Models
We set reasoning effort to medium on a sideloaded Qwen3.8 GGUF in LM Studio. The setting was accepted, never applied, and the model ran at its most verbose default. Curated downloads of the same family do apply it. A one-token test tells you which kind you have.
methodologylm-studioollamareasoningconfounds September 4, 2026The Qwen3.8-27B Buyer's Guide, Part 1: What Compression Actually Costs
Everyone recommends Qwen3.8-27B as a first local model, and nobody says which one. Fourteen packagings of the identical weights have been through this bench, spanning 15 GB to 30 GB on disk. Part 1 answers the quality question — does squeezing it smaller make it worse? — and is deliberately silent on speed, because our speed numbers are not yet comparable to each other. Part 2 fixes that.
qwenquantizationbuyers-guidelocal-llmseries