Unsloth’s latest release adds code sandboxing and trains decision models

Unsloth’s 8 October release adds code sandboxing across Windows, macOS and Linux, alongside tools for training and serving decision models. It also brings substantial changes to image workflows, model training and the Desktop browser, giving local AI users more to do without sending every task to the cloud.

Watch Desk analysis

What happened

The Unsloth release notes describe v0.1.905-beta. Users can train text or vision models as decision models with QLoRA, test them through the Decision API and export supported models for GGUF or llama.cpp. Unsloth says its sandboxing uses Bwrap on Linux, Seatbelt on macOS and MXC on Windows, with protection levels and sandbox status available in Settings.

The release also adds native support for supported ComfyUI image and video models, changes to diffusion generation, and Desktop browser improvements. The notes report up to 4.1× faster QLoRA training for Qwen3.5-35B-A3B and up to 3.3× for Qwen3-30B-A3B on specified NVIDIA hardware.

Our top picks

• Sandboxing across three operating systems
  Users can choose protection levels for code run by a model and check sandbox status in Settings.
• Train and export decision models
  The release adds QLoRA training, confidence scores for options, and export paths for supported models.
• ComfyUI models work natively
  Unsloth can recognise supported ComfyUI checkpoints and local model folders, and use their text encoders and VAEs.
• Faster training on named GPUs
  The release notes claim up to 4.1× faster QLoRA training for Qwen3.5-35B-A3B on A100 and RTX PRO 6000.
• A more useful Desktop browser
  Users can ask about open pages, confirm downloads and choose where files are saved.

Why it matters

This is a practical expansion of the local AI toolkit: model training, image generation, browsing and code execution now sit in a broader package. Sandboxing is especially relevant for anyone letting models run code, while the decision-model workflow offers a route from training to testing and serving without stitching together as many separate tools.

The performance figures are Unsloth’s claims for specified hardware, not a promise that every setup will see the same gains. And this is a beta release, so the useful question is how well the new features work in real workflows, not how many lines the changelog can fit on a screen.

Our read

There is enough here to merit more than a passing release alert. The strongest combination is practical: more control over model-run code, a clearer path for building decision models, and fewer awkward hand-offs for ComfyUI users. If you already use Unsloth, the sandbox controls and model compatibility are the first changes worth checking; if you do not, this release makes the project’s scope harder to dismiss as merely a training-speed tool.

What to watch

• How the sandbox options behave across Windows, macOS and Linux in everyday use.
• Whether the claimed training speed-ups hold across more configurations.
• Which ComfyUI models and decision-model exports receive support next.

Discussion spark: Would you trust an AI workflow more if its code ran in a sandbox, or does the quality of the controls matter more than the label?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.