Why
"Runs locally" hides two very different claims, and Qwen3.8 makes the gap unusually visible. The 27B model on a 16GB machine is the ordinary claim. The 2.4T-A95B model compressed from 4.9TB to 397GB — 91% smaller, still needing roughly 450GB of RAM — is the other one, and it is only "local" for someone who owns a server. The second thing worth pinning down is the phrase vendors use for the trade-off: "retains a lot of accuracy." Unsloth's own numbers show perplexity rising 2.58 → 4.49 and top-token agreement falling 78.9% → 66.3% between the largest and smallest 1-bit variants, in exchange for about 22% less disk. Whether that trade is worth taking is a measurement, not an opinion — and it is measurable on hardware I actually have.
How it works
The compression trick is narrower than the headline suggests. llama.cpp's IQ1_S spends 1.5625 bits per weight, of which 11 are index bits into a 2048-entry codebook. The Unsloth variants simply shrink that codebook — 1024, 512, then 256 entries — which drops the index to 10, 9 and 8 bits and the weight to 1.4375, 1.3125 and 1.1875 bpw (shipped as TQ2_0, TQ1_0 and Q1_0, names picked so the Hugging Face repo lists them at all). The claim that matters is that these are plain post-training quantizations: no quantization-aware training or distillation, which is exactly what makes them cheap to produce and worth verifying independently. The plan is to run the 27B model across several of these dtypes on hardware I own and reproduce the PPL/KLD/top-p curve myself rather than cite it, then check the practical rule of thumb — RAM+VRAM ≈ quant size, past which disk offloading quietly turns a 20 tok/s figure into something else entirely.