Unsloth ships 1-bit to BF16 GGUF quants of Qwen3.8-27B
TL;DR
- Unsloth's GGUF pack for Qwen3.8-27B spans 1-bit builds near 6 GB up to a 54.7 GB BF16 file.
- The base 27B model ships under Apache 2.0 with a vision encoder and a 262,144-token native context.
- The card claims Unsloth Dynamic V3.0 delivers over 10% better top-1% accuracy versus rival quantizers at matching size.
Unsloth has published a full ladder of GGUF quantizations for the newly released Qwen3.8-27B on Hugging Face, from 1-bit builds around 6 GB up to a 54.7 GB BF16 file. The 4-bit UD-Q4_K_M weights land between 14.3 and 17.6 GB. The model card claims Unsloth's own Dynamic V3.0 method 'Achieves >10% better top-1% accuracy compared to other quantization providers at the same size.'
The underlying model is a 27B 'Causal Language Model with Vision Encoder' with 64 layers and a native 262,144-token context that the card says is 'extensible to 1,000,000 tokens.' It ships under Apache 2.0. The repo lists 4,318,134 downloads in the past month and 81 community discussions, alongside one-line install paths for llama.cpp, Ollama, Docker and Unsloth Studio. Two researchers in our Who's Who directory posted the link within hours of the drop.
Recommended sampling splits by mode: temperature 1.0 with top-p 0.95 for Thinking Mode, and temperature 0.7 with top-p 0.80 for Instruct Mode.
Shared on Bluesky by 2 AI experts
-
We’re releasing new Qwen3.8-27B GGUFs with 10% higher accuracy. Unsloth Dynamic V3 outperforms others by >10% on Div-300, KLD & more benchmarks. We also release 1-bit quants that retain 77% accuracy. Run on 8GB RAM. B…
View on Bluesky → -
Qwen3.8-27B can now be run locally! ✨ Run on 17GB RAM via Unsloth Dynamic GGUFs. Qwen3.8-27B is by far the strongest model for its size. We also uploaded NVFP4 quants. GGUF: huggingface.co/unsloth/Qwen... Guide: unslo…
View on Bluesky →
Originally reported by huggingface.co
Read the original article →Original headline: unsloth/Qwen3.8-27B-GGUF · Hugging Face