Unsloth shipped two beta releases, v0.1.814 and v0.1.815, on September 22 and September 23 respectively, adding local support for Qwen-Image-2.1 alongside a new custom Agent Skills system and a set of infrastructure and UX improvements1,2.
Qwen-Image-2.1 can now run locally through Unsloth for both image generation and image editing. The September 22 update added image-editing capability, fixed Diffusers update issues and GGUF loading failures for the model, changed the default GGUF quantization to Q4_K_M from F16/bf16, and resolved diffusion black artifacts on A100 and consumer GPUs. The September 23 follow-up patched an httpx error affecting diffusion, restored the Fast FP8 variant of Qwen-Image-2.1 that had stopped appearing, and delivered additional diffusion and UI bug fixes.
Agent Skills and chat tooling
The releases introduce custom Agent Skills, a feature that lets users define reusable task-specific instructions for models. Skills can be imported from existing Claude Code and .agents folders and invoked inside a chat session via an @ selector. Vision models can now see images returned by MCP tools, and other accounts can connect to local model servers when the owner permits it.
Chat management received a broader overhaul: conversations are now draggable, pinnable, and movable into projects from the sidebar. The thinking UI/UX was reworked, and long reasoning blocks now render at 60 FPS, double the previous 30 FPS rate. Long code replies stream more smoothly while retaining syntax highlighting. Project names, instructions, and folders can be edited directly from the Projects page.
Training and platform changes
Training reliability improved: dataset preparation is more robust, past training runs can be resumed with correct remaining-time estimates, and dataset columns and selected splits are now mapped correctly. Hugging Face Inc. datasets work in recipes, and model uploads to the Hub exclude files from earlier exports. Downloaded models can be reused rather than fetched a second time, and CLI chats preserve loaded models' precision.
Linux distribution support expanded with in-app Debian updates and a native ARM64 .deb installer for Ubuntu 24.04 and later. A new AMD ROCm Docker image bundles Unsloth Studio and JupyterLab. Download options now span Windows, macOS, Linux x64 and ARM64 .deb packages, Linux x64 AppImage, and Windows ARM64. Accessibility additions include larger text areas and higher-contrast settings for buttons and menus, plus filterable logs with line wrapping and jump-to-latest navigation.
ANALYSIS The Agent Skills system, with its explicit Claude Code and .agents folder interoperability, positions Unsloth's desktop client as a local alternative that can absorb workflow artifacts users have already built around cloud-hosted coding agents. Pairing that with local multimodal image generation and editing via Qwen-Image-2.1 extends the tool's scope well beyond its original fine-tuning focus.