Change8

v0.1.511-beta

📦 unslothView on GitHub →
14 features🐛 16 fixes🔧 11 symbols

Summary

This release introduces support for Kimi K3 and DeepSeek v4 Flash with Unsloth Dynamic GGUFs, enables parallel chat generation, and adds a Deep Research mode. It also brings significant improvements to AMD and Intel GPU support, DoRA training, and various installer, MLX, export, and inference fixes.

Migration Steps

  1. Update Unsloth using the provided installation scripts for macOS, Linux, or WSL: `curl -fsSL https://unsloth.ai/install.sh | sh`
  2. Update Unsloth using the provided installation script for Windows: `irm https://unsloth.ai/install.ps1 | iex`
  3. If using unsupported Windows HIP GPUs, they will now fall back to Vulkan.
  4. On Linux, system ROCm libraries will be prepended to avoid bundled HIP crashes.

✨ New Features

  • Kimi K3 & DeepSeek v4 Flash can now run locally with Unsloth Dynamic GGUFs.
  • Unsloth can keep multiple chats generating in parallel, with isolated tools, uploads, self-healing, and agents.
  • New Deep Research mode plans, reads, and cites sources using your local model.
  • Improved AMD and Intel GPU support.
  • DoRA training is now available alongside LoRA and full fine-tuning.
  • Large exports can use all visible GPUs to avoid GPU 0 memory limits.
  • MLX adds streaming datasets, continued pretraining, callbacks, and better VLM and LoRA support.
  • Unsloth and saving utilities now work without bitsandbytes.
  • `unsloth start --as-subagent` allows delegation of tasks to a local Unsloth model.
  • Installer improvements for cleaner Windows installs without Winget or developer tools.
  • Vulkan devices show their real names and can be selected individually on Windows.
  • MI50 and Radeon VII support 16-bit LoRA and full fine-tuning on Linux.
  • Intel XPU enables local chat and training on Intel Arc and Data Center GPUs.
  • Release notes are available in the Unsloth update popup.

🐛 Bug Fixes

  • Fixed installer issues.
  • Fixed MLX issues.
  • Fixed export issues.
  • Fixed inference issues.
  • Fixed 4-bit NaNs, library conflicts, and long RDNA startup stalls on AMD GPUs.
  • Fixed incorrect `flash_attention_2` model output.
  • Fixed `.json` datasets and Qwen3.5/3.6 MoE and GRPO notebook setup.
  • Fixed local CLI streamed generation error handling.
  • Fixed Studio desktop reliability.
  • Fixed resume training crash recovery and MLX checkpoints.
  • Fixed loading split GGUFs from the local HF cache.
  • Fixed malformed HF token emptying the model picker's Recommended list.
  • Fixed ROCm issues on Linux by prepending system ROCm libs.
  • Extended null-kernel guard to Linux ROCm RDNA4 (gfx1201).
  • Fixed Windows HIP GPUs falling back to Vulkan.
  • Fixed `install.ps1` to use ordinal IndexOf when stripping index URL credentials.

Affected Symbols