github unslothai/unsloth v0.1.921-beta
Bug Fixes

38 minutes ago

Faster images + Restyle on Mac + faster MCP chats

This release makes image generation faster and lighter on small GPUs. Macs with Apple Silicon can now restyle videos, and chats that use MCP tools no longer reprocess the whole conversation on every message.

Highlights

  • MiniMax-H3-Image on 8 GB GPUs renders a new prompt in 36.7 s, down from 82.7 s
  • GGUF image models load with 52 to 66% less system RAM on Linux and macOS
  • Video Restyle on Mac re-renders a whole clip in a new style on Apple Silicon
  • MCP chats process 83 to 87% fewer prompt tokens in our tool tests
  • Scanned PDFs in chat are now read with OCR
  • Theme files let you share your Appearance settings

Images

  • FLUX.1 takes 16 to 28% less time per step when its weights stream from system RAM.
  • FLUX.3 klein INT8 takes 13% less time per step on T4 GPUs.
  • GTX 16xx cards no longer fail image renders with a NaN error.
  • Picking a new LoRA on an INT8 or FP8 model offers Reload and generate instead of an error.

Video

  • Cosmos3-Edge-FP8 runs about 1.45x faster on L4 GPUs.
  • Restyle on Mac needs about 42 GiB of GPU memory for bf16, or 31 GiB for q8_0.

Diffusion LoRA training

  • SDXL LoRA training takes 28 to 31% less time per step on T4 and L4.
  • Auto precision trains FLUX.2-klein 9B in about 30% less time per step on L4.
  • MiniMax-H3 re-runs reuse your encoded dataset.

Chat + MCP

  • A chat keeps its MCP tool list fixed, so each message only processes what is new.
  • Tool output length in Settings > Chat now applies to MCP results on llama.cpp, vLLM, Ollama and Custom connections.
  • Attached Word files no longer lose moved paragraphs, date fields or furigana text.
  • Imported Claude Code and Codex chats keep their images.
  • A reasoning model that runs out of tokens while thinking now shows advice and Retry.

Desktop + Mac

  • Export or import themes in Settings > Appearance > Theme file.
  • Unsloth Desktop no longer gets stuck on an incomplete install on some Windows AMD PCs.
  • Macs with a chat-only GGUF install no longer show small models as "Does not fit".
  • A lost reply during a session refresh no longer signs you out of every tab.

Training

  • AsyncGRPO steps are 1.46x to 2.47x faster on A100 with TRL 1.17 or newer.
  • On AMD ROCm, Qwen3.5 training packs sequences by default, about 1.4x the tokens per second on Strix Halo.
  • DoRA training is about 3 to 5% faster per step.
  • Gemma-4 LoRA training no longer fails on mixed-length batches with PyTorch older than 2.14.
  • Dates in JSON and JSONL datasets no longer gain a 00:00:00 time.
  • PEFT 0.22 works, and DPO with a PEFT model no longer crashes.

Don't miss a new unsloth release

NewReleases is sending notifications on new releases.