Compressing 27 billion parameters again
Unsloth, which makes local LLM quantization tools, released a new Qwen3.8-27B GGUF through its official account. According to the official documentation and its Hugging Face repository, the new quantization method, Dynamic v3.0, preserves more model quality than the earlier v2.0 while keeping file size the same. It says it recorded more than 10% higher accuracy than other providers' quantizations on several benchmarks, including Div-300 and KL Divergence.
This isn't the first announcement. Qwen3.8-27B was released as an open-weights model on August 15, and right after that Unsloth put out a quantized version that could run locally on 17GB of RAM. Dynamic v3.0 is the follow-up improvement.
Even at 1 bit, 77% remains
The most striking figure is the 1-bit quantization. Pushing quantization bits to the extreme usually causes accuracy to drop sharply, but Unsloth says Dynamic v3.0's 1-bit quantization keeps 77% of the original's accuracy. This version can run on 8GB of RAM, which means you could run a 27-billion-parameter-class model on laptop RAM alone, without a consumer graphics card.
Of course, 77% accuracy also means a performance loss compared with the original. For work where precision matters, it's safer to choose a higher quantization level than 1 bit, and running on 8GB of RAM should be seen as an option for very low-spec environments only.
Use it right away in llama.cpp
The new GGUF is compatible with most inference engines, including llama.cpp, and is fully supported in Unsloth Studio and Desktop. Download the quantization level (bit width) you want from the Hugging Face repository and plug it straight into your existing local run pipeline.
However, some users commented on X that "the Dynamic v3.0 GGUF isn't fully compatible with standard llama.cpp builds and is more stable in Unsloth's own runtime." Actual compatibility may vary with the llama.cpp build you use, so test it in your own environment before adopting it.
Running without a GPU is not the same as running fast
That it can run on 8GB of RAM is separate from whether it runs fast enough for real use. Unsloth published accuracy figures but didn't separately share token generation speed on low-spec hardware. Before putting it to work, it's safer to check response speed on your own hardware.