Unsloth

Quantization and fine-tuning house (Daniel Han) — 2-5x faster LoRA/QLoRA fine-tuning at lower VRAM, and high-quality GGUF releases. Most GGUFs on the doombox are unsloth builds, identifiable by the UD- (Unsloth Dynamic) prefix on quants: Qwen3.8-27B-UD-Q4_K_M, Qwen3.6-35B-A3B-UD-Q3_K_XL, gemma-4-12B-it-qat-UD-Q4_K_XL — dynamic per-layer quant allocation rather than one flat scheme.

Local footprint (llama-swap, 3090): Qwen3.8-27B Q4_K_M (the most-referenced file, 14 entries), Qwen3.6-27B MTP, Qwen3.6-35B-A3B MoE MTP, gemma-4-12B-it-qat (Google’s QAT build, unsloth GGUF).

Links: Qwen model family (whose weights they ship) · LoRA (their training specialty) · Gestalt Labs/OrcaRouter (other labs in the local rotation)