Unsloth has pushed a batch of fixes to Studio and its MTP (Multi-Token Prediction) layer that are worth pulling immediately if you run local inference or fine-tuning workflows.
The headline improvement is MTP performance. Previous releases left speed gains unrealized on Macs, CPUs, and GPUs. That's now fixed — the release notes call the improvement "much better." The underlying change was emitting a single comma-chained --spec-type flag for the CPU/Mac MTP path, and enabling draft-MTP on vision MTP GGUFs by dropping an incorrect vision gate that was blocking it.
On the Studio side, two friction points that could silently break a session are resolved. First, users were getting stuck on the reset-password page — a stale must_change_password flag in localStorage was the culprit, and it's now reconciled properly on load. Second, the auto-load cascade was allowed to run unbounded; it's now capped to prevent runaway attempts that could lock up the UI.
The desktop shortcut breaking after an update was another silent killer for daily workflows. The fix regenerates the desktop launcher on unsloth studio update across macOS, Linux, and Windows — so the shortcut you pinned stays valid after each update.
Offline mode gets broader support in this release. The loader now honors HF_HUB_OFFLINE and TRANSFORMERS_OFFLINE environment variables, which matters if you're running in air-gapped or bandwidth-constrained environments and don't want surprise network calls.
Smaller but still meaningful: the unsloth studio update command itself was broken and is now fixed, toast close-button clicks and light-mode hover states are repaired in the frontend, and the loss function patch for Qwen3.5 models is back in place after a regression. GGUF multi-image chat handling also received a fix.
Run the update. On macOS, Linux, or WSL:
curl -fsSL https://unsloth.ai/install.sh | sh
On Windows:
irm https://unsloth.ai/install.ps1 | iex
If you use vision models with MTP GGUFs, test inference speed before and after — the draft-MTP path is now active where it previously wasn't. If you work offline or in restricted environments, set HF_HUB_OFFLINE=1 and verify the loader respects it. And if you run Qwen3.5, confirm your loss function is patching correctly post-update.