Unsloth ships Qwen3.8-27B GGUFs with Dynamic V3.1 quantization
Start here
Unsloth released updated Qwen3.8-27B GGUF quantizations built with its Unsloth Dynamic V3.1 (preview) algorithm. Unsloth says the new quants gain 5-10% accuracy at the same size on most benchmarks, including Terminal Bench 2.1, DeepSWE v1.1, LiveCodeBench v6, GPQA Diamond, MMLU-Pro and AIME25. Most quants from 2-bit to 6-bit were updated, but the 3-bit UD3 version was left unchanged because Unsloth found it performed on par with UD3.1. NVFP4 quants currently run only in vLLM, not SGLang, and require NVIDIA Blackwell GPUs such as RTX 50-series, DGX Spark, B200 and B300. Unsloth also disputes GSQ-RCO's benchmark claims about ISTA quants, saying it ran 14 trials versus an estimated one and reached the opposite conclusion. Qwen3.8-27B is a 27B dense model with vision, reasoning and a 256K context window that runs on 17GB RAM/VRAM setups.
Generated from reports · updated 37 min ago
Report timeline
Follow the reports to see the event from different sides.
- Hacker News · AIUnsloth releases Qwen3.8-27B GGUFs with Dynamic V3.1 quantization update
Unsloth released updated Qwen3.8-27B GGUFs built with its Unsloth Dynamic V3.1 (preview) quantization algorithm, which it says improves accuracy by 5-10% at comparable sizes across most benchmarks, including Terminal Bench 2.1, DeepSWE v1.1, LiveCodeBench v6, GPQA Diamond, MMLU-Pro and AIME25. Quants from 2-bit to 6-bit were updated while 3-bit UD3 was left unchanged because it performed on par with UD3.1. The post also disputes GSQ-RCO's benchmark claims about ISTA quants, saying Unsloth ran 14 trials versus an estimated 1 and reached the opposite conclusion, and notes Qwen3.8-27B is a 27B dense model with vision, reasoning and a 256K context window that runs on 17GB RAM/VRAM setups.
Heat for this event
Not enough continuous observations yet to draw a trend.