●Change8

v0.1.51-beta

📦 unslothView on GitHub →
✨ 11 features🐛 27 fixes🔧 18 symbols

Summary

This release introduces Kimi K3 local execution, parallel chat capabilities, and 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, WSL, or Windows.

✨ New Features

  • Kimi K3 model support for local execution with Unsloth Dynamic GGUFs.
  • Parallel chat functionality allowing multiple conversations to generate simultaneously.
  • Deep Research mode for local models to plan, read, and cite sources.
  • Improved AMD GPU support including RDNA2, Radeon, Ryzen, Strix Halo, and workstation GPUs.
  • DoRA training support alongside LoRA and full fine-tuning.
  • Intel XPU support for local chat and training on Intel Arc and Data Center GPUs.
  • Large export functionality utilizing all visible GPUs to overcome GPU 0 memory limits.
  • MLX enhancements including streaming datasets, continued pretraining, callbacks, and better VLM and LoRA support.
  • Installer improvements for cleaner Windows installations without Winget or developer tools.
  • Ability to delegate tasks to a local Unsloth model using `unsloth start --as-subagent` for Codex, Claude Code, and Pi.
  • Optional full-page grounding in Deep Research to read top search results into temporary RAG before synthesis.

🐛 Bug Fixes

  • Fixed 4-bit NaNs, library conflicts, and long RDNA startup stalls for AMD GPUs.
  • Unsupported Windows HIP GPUs can now fall back to Vulkan.
  • Vulkan devices now show their real names and can be selected individually.
  • Fixed incorrect `flash_attention_2` model output.
  • Unsloth and saving utilities now work without bitsandbytes.
  • Fixed `.json` datasets and Qwen3.5/3.6 MoE and GRPO notebook setup.
  • Studio: Fixed malformed HF token emptying the model picker's Recommended list.
  • Studio: Fixed local CLI streamed generation error handling.
  • Studio: Fixed tab navigation responsiveness.
  • Studio: Fixed desktop reliability.
  • Studio: Fixed resume training crash recovery and MLX checkpoints.
  • Studio: Fixed loading split GGUFs from the local HF cache.
  • Studio: Fixed HF token tick display only after validation.
  • Installer: Enabled ROCm torch on RDNA2 (gfx1030-1036) on Windows.
  • Fixed prepending system ROCm libs on native Linux to avoid bundled HIP crash.
  • Studio: Extended the _grouped_mm null-kernel guard to Linux ROCm RDNA4 (gfx1201).
  • Recognized Radeon 8065S (Gorgon Halo / Ryzen AI Max 400) as gfx1151.
  • Fixed ordinal IndexOf when stripping index URL credentials in install.ps1.
  • Installer: Routed Strix to the AMD arch index on rocm7.2 and PCI detection hint.
  • Studio: Moved sidebar search into the header.
  • Studio: Reused MLX prompt cache across turns instead of re-prefilling.
  • Unsloth start improvements: download progress, server reuse, and safe model switching.
  • Baseline two benign unsloth-zoo test-file findings in scan_packages.
  • Studio: Show system-wide VRAM in the multi-GPU System tab view on ROCm.
  • Studio: Classified embedding models from the HF cache and honored offline mode.
  • Unsloth start: Added local subagents for Claude Code, Codex, OpenCode and Pi.
  • Installer: Routed Strix to AMD gfx index on ROCm 7.14.

Affected Symbols

unslothflash_attention_2bitsandbytesQwen3.5Qwen3.6MLXROCmHIPVulkanGGUFLoRADoRAUD-IQ1_SUD-Q4_K_XLUD-Q8_K_XLKimi K3unsloth start --as-subagentUNSLOTH_RESEARCH_AUTO_SCRAPE