v0.1.526-beta
📦 unslothView on GitHub →
✨ 12 features🐛 34 fixes🔧 8 symbols
Summary
Unsloth now supports running Kimi K3 and DeepSeek v4 Flash models locally with Dynamic GGUFs, featuring improved download efficiency and DSpark support for faster inference. This update also includes numerous bug fixes, installation improvements, and enhanced studio features.
Migration Steps
- Update installation scripts for macOS, Linux, and WSL using: `curl -fsSL https://unsloth.ai/install.sh | sh`
- Update installation scripts for Windows using: `irm https://unsloth.ai/install.ps1 | iex`
✨ New Features
- Kimi K3 and DeepSeek v4 Flash models can now run locally with Unsloth Dynamic GGUFs.
- Added more efficient and faster downloading for Colab, low memory systems, and high memory/CPU systems.
- Auto fallback to HTTP if XET is stuck.
- DeepSeek V4 Flash models automatically detect multi-GPU setups and can offload model layers to system memory.
- DSpark support is now default for DeepSeek V4 Flash, providing 2x faster inference.
- Kimi K3 model supports low, high, and max reasoning efforts.
- Kimi K3 models automatically detect multi-GPU setups and offload layers automatically.
- Studio: Report active training status.
- Studio: Apply saved per-model settings on API loads, add API monitor.
- Offline: Detect an unreachable hub, not just dead DNS.
- Studio: Charge the split compute-buffer rate on a multi-GPU auto fit.
- Studio: Complete translation coverage for all supported locales.
🐛 Bug Fixes
- Many bug fixes and Mac fixes.
- Smoother installations.
- Fix two pre-existing Backend CI failures blocking every open PR.
- Warn loudly when requested packing is silently disabled.
- Preserve reasoning across turns for Qwen3.x.
- Make tests/utils runnable again and cover test_packing.py in CI.
- Uninstaller: reject unrecognized arguments instead of uninstalling.
- Make the uninstaller test suites run on macOS and Windows PowerShell 5.1.
- Studio: Don't disarm Ctrl+C in agents launched by unsloth start.
- Studio: keep a slow model load alive through a proxy timeout.
- Tighten the comments around the uninstaller argument guard.
- Route the explicit Vulkan setup failures through the setup failure helpers.
- Honor an explicit map_eos_token=False in get_chat_template.
- Fix desktop repair version skew and output decoding.
- Desktop: Improve the installation message.
- Studio: make the API keys list scrollable.
- Desktop: notarize the final macOS disk image on release.
- Desktop: preserve first app window size after setup.
- data-recipe: recognize current sampling progress log line.
- Desktop: clean up the backend on Unix termination signals.
- Desktop: show the app version in Settings > About.
- Studio: Fix hub search input overlapping the format filter at narrow widths.
- Desktop: exit the backend when the app that spawned it dies.
- Desktop: open uploaded files via the OS instead of window.open.
- Run OpenClaw in the launched project workspace.
- Desktop: drop documents into a chat instead of rejecting them as non-GGUF.
- Fix multiline prompts through Windows npm shims.
- docs: fix `--quantization` flag in unsloth-cli.py usage example.
- Studio: Fix GGUF classification for custom MTP scan roots.
- Guard VLM detection against config.architectures being None in the save paths.
- Centralize device backend helpers.
- Make the auto-appended EOS token deterministic.
- Fix desktop repair version skew and output decoding.
- Assert the Vulkan installer bailouts fail closed via the setup failure helper.