Unsloth adds desktop command palette and lower-memory LoRA training

Unsloth’s October 1, 2026 release adds a desktop command palette and shareable GGUF run settings; the settings can be shared without loading models. The update also improves desktop controls, image loading and model-fit warnings, and lets users continue responses and fix HTML canvas errors in chat. Laya decision models respond up to 4.1x faster, and hosted providers are available through Connections. For 4-bit LoRA training, pre-quantized checkpoints use their original packed weights, reducing memory use while preserving accuracy; the support covers NVFP4, INT4 and MXFP4 checkpoints. Unsloth gives Qwen3.8-27B-NVFP4 as an example, at 40.2 GB peak memory rather than 72.9 GB. The release also adds FastModel fine-tuning support for T5, T5Gemma, BART and Marian, and adds unsloth eval for checkpoints and LoRA adapters.