Unsloth's Dynamic 3.0 makes Qwen3.8-27B smaller and more accurate at once

Unsloth's Dynamic 3.0 quantization, released on 19 August for Qwen3.8-27B, does something quantization usually forces you to trade off: it makes the model files smaller and closer to the full-precision original at the same time. Co-founder Daniel Han put it plainly in the release thread. "We're releasing new Qwen3.8-27B GGUFs with 10% higher accuracy for the same size."

The method is post-training, not retraining. Unsloth's own page is explicit that "we do NOT use QAT or QAD. Everything is done through post-training quantization." What changed in 3.0 is how precision is spread across the model. Every layer is quantized to its own bit width, guided by an importance matrix that flags which weights carry the most signal. Version 3.0 pushes more of the low-importance weights down harder. One user who inspected the new 3-bit build counted 24 two-bit tensors, about 7% of the model, where the previous build had none. That freed budget is why the files shrank. A common file dropped from 8.39 GB to 6.77, a 19% cut, with the saved bits reinvested where they matter.

Unsloth measures quality with a 300-prompt test that compares each quant's next-token choices against the full model across coding, math and long-document prompts. It reports the new files beating both its old version and rival providers at matched size. An independent check in the thread, on the mid-size file, measured a KL divergence of 0.019 from the base model and 94% top-token agreement at 13 GB, and called it the best value for a 16 GB card.

The limit is at the bottom. Unsloth's page says the smallest 1-bit files fall below 10% on that same test and "should not be used for agentic use-cases."

Key takeaway: for Qwen3.8-27B on a 16 to 24 GB card, the Dynamic 3.0 build is a real upgrade. Pull the Q3 or Q4 XL file and you get a smaller download that tracks the full model more closely than the previous version did. Keep the 1-bit files for simple lookups, not tool calling.

Sources: Unsloth Dynamic v3 announcement (discussion #74) · Unsloth Dynamic 3.0 methodology page · The previous (14 August) version, commit 408fcc1807ab