Notes from the lab.
Low-bit models think longer
Squeezing a reasoning model into less memory can make it second-guess more. Measured for 2-bit weights and 8-, 4- and 2-bit KV caches, against a published decode-time penalty, and with AZ-Focus.
Read →14 October 2026More correct answers within your budget
Scored the way people run reasoning models, with a time or token budget, AZ-Focus answers more questions correctly and finishes them: GLM-5.3-Flash on two Sparks, Qwen3.8-Flash-Next on one, and a 2-bit GLM quant.
Read →11 October 2026Half the thinking, same answers: Qwen3.8-Flash-Next on one DGX Spark
Qwen3.8-Flash-Next-AZ-Focus against the base model on one Spark with Mia's TensorFold recipe: 46% fewer thinking tokens across 1,298 answers, the same accuracy, and a 60-question set finished 3.3× sooner.
Read →8 October 2026Same speed, half the wait: GLM-5.3-Flash on two DGX Sparks
GLM-5.3-Flash-AZ-Focus against stock on identical software: the same tokens per second, about half the tokens, and answers finished in about half the time.
Read →