●Change8

v0.1.47-beta

📦 unslothView on GitHub →
✨ 19 features🐛 12 fixes🔧 4 symbols

Summary

This release introduces significant enhancements to Unsloth Studio, including 3x longer context lengths, support for GLM 5.2 GGUFs, and new parallel operational capabilities. Key new features focus on chat experience (forking, queueing) and secure cloud access via Cloudflare.

Migration Steps

  1. Use `unsloth studio --secure` to enable secure HTTPS global access via Cloudflare.
  2. Update Unsloth Studio using the provided installation scripts (curl for MacOS/Linux/WSL, irm for Windows) to ensure version `2026.6.8` or `v0.1.47-beta` is installed.

✨ New Features

  • Support for GLM 5.2 GGUFs in Unsloth Studio.
  • Achieved 3x longer context lengths via a new auto-fit algorithm with MTP.
  • Introduction of bypass permissions mode in Unsloth Studio.
  • Forkable chats functionality added to the chat canvas.
  • Queue-able chats allow submitting new prompts while generation is running.
  • New hub for model discovery with trending feed, search, and custom model paths.
  • Parallel modules support, allowing simultaneous operation of Export, Chat, Training, and Recipes.
  • HTTPS Cloudflare support added; use `unsloth studio --secure` for secure global access.
  • Chat canvas now supports inline HTML canvas cards, Code view, and visible raw code for DiffusionGemma.
  • Chat search now covers all messages and prioritizes user messages.
  • Hub now defaults downloads to the faster Xet transport with HTTP fallback.
  • New "Load on selection" toggle in the Hub to configure load options before loading a model.
  • Live image formation display for DiffusionGemma during denoising, with accurate speed stats.
  • New `--secure` Cloudflare-only mode for end-to-end encrypted studios.
  • New API server monitor added to Studio.
  • Faster API calling and reduced latency.
  • Streamlined logs showing throughput and latency, with removal of bloated logs.
  • Improved support for Blackwell RTX 50X and 60X GPUs.
  • Installer now auto-repairs broken or CPU-only PyTorch installs and warns on silent CPU fallback.

🐛 Bug Fixes

  • Fixed fast inference speed-ups on the latest vLLM (0.22+).
  • Tensor parallelism (MTP) now recovers on failure instead of crashing.
  • Fixed DeepSeek-OCR and other vision models loading/running errors.
  • Fixed silent downgrading to CPU instead of GPU.
  • Fixed chat model VRAM not being freed during training when GPU is tight (now only frees when necessary).
  • Studio now steps through a recovery ladder if llama-server hard-crashes at startup.
  • Improved GRPO training reliability with vLLM.
  • Training startup made more reliable with clearer errors for invalid VLM batches.
  • Studio now cleans up leftover backend processes more reliably after crashes, restarts, or interrupted shutdowns.
  • Fixed Responses tool output content arrays.
  • Fixed training output directory escaping outputs root for models on another drive.
  • Fixed llama.cpp prebuilt skipping already-installed same-release fallback.

Affected Symbols

DeepSeek-OCRDiffusionGemmavLLM (0.22+)llama-server