Unsloth's Dynamic v3.0 Quantization Brings Qwen3.8-27B to 8GB Machines
Unsloth (Daniel Han) announced Dynamic v3.0 dynamic quantization on Aug 19 and released corresponding Qwen3.8-27B GGUF models on Aug 20. The team claims 10%+ higher Top-1% accuracy at equal model size versus other quantization providers, and similar leads on Divergence-300 and KL Divergence benchmarks. Extreme 1-bit/2-bit variants (UD-IQ1M / UD-IQ2S) retain 77% accuracy while shrinking the 27B model to 6-8GB, runnable on an 8GB Mac.
Confirmed
- Daniel Han posted Dynamic v3.0 on Aug 19; Qwen3.8-27B GGUF quants followed on Aug 20
- Official numbers: 10%+ Top-1% accuracy gain at equal size; 10%+ lead on Div-300 and KLD benchmarks
- 1-bit/2-bit variants keep 77% accuracy at 6-8GB, running with only 8GB memory
Why it matters
- Per @Arindam1729, the core idea is keeping the most important weights at high precision while aggressively quantizing insensitive ones, rather than uniform low-bit quantization, halving size with better quality
- It targets llama.cpp local deployment, balancing memory footprint and quality
- 1-bit/2-bit options bring a 27B-class model to 8GB devices, sharply lowering local deployment barriers
2026-08-19 ~ 2026-08-21 · 5 related posts
Primary sources
- Unsloth releases Dynamic v3: Qwen 27B quantization gains 10% accuracy — danielhanchen ·
- Unsloth Releases Qwen3.8-27B GGUFs with 10% Higher Accuracy — danielhanchen ·
- Unsloth Releases Dynamic 3.0 GGUF: Smaller Size, Better Quality — Arindam_1729 ·
- [source] Unsloth releases Dynamic v3: Qwen 27B quantization gains 10% accuracy — danielhanchen · 2026-08-19
- [source] Unsloth Releases Qwen3.8-27B GGUFs with 10% Higher Accuracy — danielhanchen · 2026-08-20
- Qwen 27B now runs on 8GB Mac via Unsloth's extreme quantization — SumitGup · 2026-08-20
- [source] Unsloth Releases Dynamic 3.0 GGUF: Smaller Size, Better Quality — Arindam_1729 · 2026-08-20
1 near-duplicate retellings: pcuenq