Piotr Migdał tested what happens when you compress Qwen3.8 27B enough to run on a consumer graphics card. His August 26 study put Unsloth's GGUF model versions through llama.cpp on GPQA Diamond, IFBench and Terminal-Bench 2.1.
The 4-bit Q4_K_M file weighs about 17 GB, compared with roughly 55 GB for the full BF16 model. In Quesma's 89-task Terminal-Bench 2.1 run, the compressed model matched BF16. Quesma says it fits the 24 GB memory of an RTX 4090 with room for context.
Further compression had a cost. The 10.7 GB, 2-bit UD-Q2_K_XL lost some coding performance, even though the instruction-following test remained competitive. At 1 bit, UD-IQ1_S fell near random guessing on GPQA Diamond; longer reasoning did not recover the lost accuracy.
Quesma ran the tests on rented GPUs, with a fixed F16 key-value cache. Some Unsloth files used in the comparison were replaced after the runs, the author notes, so reproducing the scores requires the original versions.