How much accuracy does the 1-bit Unsloth quant keep?

Unsloth reports its UD-IQ1_S 1-bit build of Qwen3.8-27B holds around 72% top-1% accuracy while shrinking the file 89% to 6.2 GB. The mid-range UD-Q2_K_XL at 9.83 GB lands roughly 8% ahead of the next best provider on top-1%, and Unsloth notes that build wrote a working HTML program where earlier quants broke.

Which inference engines run Dynamic 3.0 GGUFs?

Dynamic 3.0 files use the standard GGUF container, so Unsloth says they work with most inference engines, including llama.cpp and Unsloth Desktop. Unsloth publishes them on Hugging Face under unsloth/Qwen3.8-27B-GGUF, which means any llama.cpp-compatible runner can download and load a quant without extra tooling.

How is Dynamic 3.0 different from Dynamic 2.0?

Dynamic 3.0 swaps the Dynamic 2.0 calibration data for a broader, higher-quality imatrix set drawn from agentic coding, chat and multilingual sources. Unsloth also made the per-layer choice of quantization type finer-grained and added more quantization techniques to pick from, so a file of a given size carries more of the original model's behaviour.

What does the Divergence-300 @32 score measure?

Divergence-300 @32 is Unsloth's own drift test: it runs greedy argmax decoding for 32 tokens on 300 held-out examples and compares each quant's output against the BF16 original. Unsloth draws the held-out text from sources such as DeepSWE and Terminal Bench so the score reflects unseen data rather than the calibration set.

Are the smallest Dynamic 3.0 files missing anything?

Yes. Unsloth removed the multi-token prediction (MTP) module from every Dynamic 3.0 quant of Qwen3.8-27B at 8.37 GB and below, which frees about 500 MB of disk space. Larger builds keep MTP, and Unsloth stores those files in a separate MTP folder inside the Hugging Face repository.

Unsloth · 2026-08-19 · major

Unsloth Dynamic 3.0 — GGUF quants keep 10% more accuracy per gigabyte

Unsloth Dynamic 3.0 is a new GGUF quantization recipe that Unsloth says keeps over 10% more top-1% accuracy at the same file size than any other provider. The first full v3.0 set covers Qwen3.8-27B.

Hugging Face model card banner for the Unsloth Qwen3.8-27B GGUF quantizations

Unsloth's third-generation quantization recipe squeezes more accuracy out of the same GGUF file size.

Key specs

Accuracy gain	>10% top-1% at same size
Smallest quant	6.2 GB

Quick facts

Maker	Unsloth
Method	Post-training quantization (no QAT or QAD)
First model covered	Qwen3.8-27B
Size range	6.2 GB (UD-IQ1_S) to 54.7 GB (BF16)
Runs on	llama.cpp, Unsloth Desktop
Quality metrics	KL Divergence, Divergence-300 @32
Where to get it	Hugging Face — unsloth/Qwen3.8-27B-GGUF

What is it?

Dynamic 3.0 changes how Unsloth turns full-precision LLM weights into GGUF files, and Unsloth reports the result holds over 10% more top-1% accuracy than any other provider's file of the same size. The first complete v3.0 set covers Qwen3.8-27B, spanning eleven builds from a 6.2 GB 1-bit quant up to 54.7 GB BF16. Everything ships on Hugging Face and loads in ordinary llama.cpp-compatible runners.

How does it work?

Three changes drive the gain: a larger, higher-quality imatrix calibration set drawn from agentic coding, chat and multilingual text; finer per-layer decisions about which quantization type each layer receives; and a wider menu of quantization techniques to choose from. Unsloth states it does not train on the calibration data and uses neither quantization-aware training nor distillation — Dynamic 3.0 is purely post-training. Quality is then checked with KL Divergence and Unsloth's held-out Divergence-300 @32 test.

Why does it matter?

File size is the binding constraint for anyone running a large open model on a single GPU or a laptop, so accuracy-per-gigabyte decides which models are usable at all. A Dynamic 3.0 quant can do work that previously needed the next size up, which frees VRAM for longer context or a bigger batch. Unsloth's Qwen3.8 GGUFs were downloaded 5.1 million times in five days, so the recipe reaches a large share of local-LLM users immediately.

Who is it for?

people running open models locally

Frequently asked questions

How much accuracy does the 1-bit Unsloth quant keep?: Unsloth reports its UD-IQ1_S 1-bit build of Qwen3.8-27B holds around 72% top-1% accuracy while shrinking the file 89% to 6.2 GB. The mid-range UD-Q2_K_XL at 9.83 GB lands roughly 8% ahead of the next best provider on top-1%, and Unsloth notes that build wrote a working HTML program where earlier quants broke.
Which inference engines run Dynamic 3.0 GGUFs?: Dynamic 3.0 files use the standard GGUF container, so Unsloth says they work with most inference engines, including llama.cpp and Unsloth Desktop. Unsloth publishes them on Hugging Face under unsloth/Qwen3.8-27B-GGUF, which means any llama.cpp-compatible runner can download and load a quant without extra tooling.
How is Dynamic 3.0 different from Dynamic 2.0?: Dynamic 3.0 swaps the Dynamic 2.0 calibration data for a broader, higher-quality imatrix set drawn from agentic coding, chat and multilingual sources. Unsloth also made the per-layer choice of quantization type finer-grained and added more quantization techniques to pick from, so a file of a given size carries more of the original model's behaviour.
What does the Divergence-300 @32 score measure?: Divergence-300 @32 is Unsloth's own drift test: it runs greedy argmax decoding for 32 tokens on 300 held-out examples and compares each quant's output against the BF16 original. Unsloth draws the held-out text from sources such as DeepSWE and Terminal Bench so the score reflects unseen data rather than the calibration set.
Are the smallest Dynamic 3.0 files missing anything?: Yes. Unsloth removed the multi-token prediction (MTP) module from every Dynamic 3.0 quant of Qwen3.8-27B at 8.37 GB and below, which frees about 500 MB of disk space. Larger builds keep MTP, and Unsloth stores those files in a separate MTP folder inside the Hugging Face repository.

Try it

https://huggingface.co/unsloth/Qwen3.8-27B-GGUF