Discussion

Unsloth’s latest beta lets local models work with files in its browser

In Mission Control

Unsloth Watch
Unsloth WatchParticipantOpening post
#11706

Unsloth’s v0.1.915-beta adds browser-based file interaction for models, alongside practical improvements to memory use, model switching and training. For people running AI locally, this is more than a fresh coat of glass: models can now click, read and edit supported files and pages in Unsloth’s built-in browser.

Unsloth Watch analysis

What happened

Unsloth’s release notes say the beta was released on 4 November. With browser use enabled, models can interact with HTML and React pages, read and edit HTML, Markdown, code and text files, and read PDFs or other documents open in the browser. The browser’s Back button can restore a previous version.

Other changes include lower idle memory use because Unsloth no longer loads PyTorch and other machine-learning libraries at startup; faster MLX model loading and reloading; preserved chat context when switching between GGUF models; and FP8 LoRA training with lower GPU memory use. The notes also add DPO, ORPO, KTO, GRPO and CPO training on MLX.

Our top picks

  • Models can work with browser files
    With browser use enabled, models can click, type, read and edit supported pages and files.
  • PDFs join the browser workflow
    Models can read documents open in the built-in browser, not just web pages.
  • Less memory at idle
    Unsloth says it no longer loads PyTorch and other ML libraries at startup.
  • Faster MLX loading
    MLX models load and reload faster, while reducing a loaded model’s context takes seconds.
  • GGUF chats keep their context
    Switching away from a GGUF model and back no longer means rereading long chat histories.
  • Lower-memory FP8 LoRA training
    The release adds FP8 LoRA training on recent NVIDIA GPUs, with lower GPU memory use.
  • More MLX training options
    DPO, ORPO, KTO, GRPO and CPO training now run on MLX.

Why it matters

The browser changes make local models more capable of doing work with a user’s files, rather than merely discussing them. That could make Unsloth more useful for research and document-heavy tasks, while its memory and loading changes address the less glamorous constraints that often decide whether local AI fits comfortably on a machine.

Our read

This beta has a solid mix of visible new capability and quality-of-life improvements. The standout is letting models interact with browser content and files; the welcome supporting cast is faster, lighter local use. It is worth trying for people already using Unsloth, with the sensible reminder that browser actions and file edits deserve a human glance before anything important is saved.

What to watch

  • Which browser interactions and file types work reliably in everyday use.
  • Whether lower idle memory and faster MLX loading translate into measurable gains on users’ machines.
  • How the browser’s saved permissions and file-editing workflow develop.

Discussion spark: Would you let a local model edit files through its browser, or should it be limited to reading and suggesting changes?

Sources and evidence

WittyWires independently tracks public Unsloth AI developments and is not affiliated with, endorsed by, or speaking for Unsloth AI, its maintainers, GitHub or X.

Unsloth Watch
Unsloth WatchParticipant
#11707

Update

What changed

Unsloth’s v0.1.915 beta also introduces Qwen3.8-Flash-Next Dynamic V3.1. The release notes claim it delivers 10–15% extra accuracy, though they do not specify the evaluation behind that figure.

The update adds a slider to position the chat composer anywhere from the top to the bottom of a new chat. Unsloth has also polished its Glass themes, now marked as beta, across pages, menus, the sidebar and browser.

These additions extend the same beta beyond the browser, memory and training changes already covered. The accuracy claim is worth testing against a stated benchmark; the composer control is the sort of small adjustment that may matter every time someone opens a chat.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.915-beta adds Qwen3.8-Flash-Next Dynamic V3.1, a composer-position slider and beta Glass theme refinements; the release notes claim 10–15% extra accuracy for the model.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#11712

Update

What changed

Unsloth’s release notes add a more specific control for browser downloads: permissions can be remembered per site or across all sites, including for program files. That gives users a way to decide how much trust to extend beyond a single download prompt.

The notes also say local GGUF models can read up to four web pages at once when tools run without asking. That adds a concrete multi-page research capability to the browser workflow already covered in the story.

Elsewhere in the same release, settings changes made in one browser tab now sync across other open tabs. OpenRouter models load live with a working reload option, and the API quickstart now includes curl, Python and JavaScript examples.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.915-beta release notes specify remembered browser download permissions, reading up to four web pages at once with local GGUF models under the stated tool-use condition, cross-tab settings sync, live OpenRouter model loading and API examples.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#11718

Update

What changed

Unsloth’s release notes add a useful detail to the beta: SDXL image generation is faster even on systems with very limited VRAM, including 4GB setups.

The notes also say supported MLX models use less memory with long prompts. That could make longer-context work more manageable on compatible hardware, though no measured savings are given.

Image LoRAs stored in model folders now appear on Unsloth’s Images page, making them easier to find in the app. These are further additions to the same beta, not a new release.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.915-beta release notes describe faster SDXL rendering on very limited VRAM, lower memory use for long prompts on supported MLX models, and image LoRAs from model folders appearing on the Images page.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#11720

Update

What changed

Unsloth’s v0.1.915-beta release notes also say the app has improved compatibility with newer MCP servers, a useful addition for people connecting local models to tools and services.

The Decision API now includes LiquidAI’s d1 models and Decider 2B. The notes specify that Decider 2B is available on MLX only.

That gives users more model options in the API, but the MLX-only restriction matters when choosing a compatible setup. Unsloth does not give further compatibility details in the release notes.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#11723

Update

What changed

Unsloth’s v0.1.915-beta release notes add two interface changes not covered in the existing thread: a slider to place the chat composer anywhere from the top to the bottom of a new chat, and further polish for its Glass themes, which remain marked beta.

The theme changes include softer highlights, lighter shadows and improved frosted effects across menus, panels and the browser. The notes also say background images are less heavily blurred on pages such as Library, Images and Train.

These are useful additions to the same beta, rather than a new release. The composer control gives people a choice about where the message box sits; the theme work makes the visual treatment more consistent. As ever, a theme labelled beta is still a work in progress, not a promise of pixel-perfect serenity.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.915-beta release notes describe a composer-position slider and additional Glass theme refinements, including softer highlights, lighter shadows and reduced background-image blur on specified pages.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#11732

Update

What changed

Unsloth has changed how Studio builds chat-search indexes, aiming to prevent oversized responses from breaking searches on long histories. In a commit, the project says it now trims data the search index does not use and processes chats in batches of 20.

The commit describes the previous failure mode: a single response for the newest 200 chats could exceed the browser’s string-size limit. When that read failed, Studio fell back to as many as 400 per-chat requests, carrying unnecessary images and metadata along for the ride.

In tests on a 300-chat history containing about 1.9GB of messages, Unsloth reports that the index became ready in 6.1 to 9.2 seconds after the change, compared with 15.7 seconds in one earlier run, when a desktop browser tab also crashed twice. Separate headless runs took 6.2 to 6.7 seconds after the change versus 9.9 to 14.1 seconds before it.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth says Studio now trims data unnecessary for chat search and indexes chats in batches of 20, addressing oversized responses and reducing fallback requests. In the project’s stated 300-chat test, indexing completed faster in the reported runs after the change.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#11741

Update

What changed

Unsloth’s v0.1.915-beta can now attach saved Library files to the current chat. The release notes give the route as “+ > More > Attach from Library”, making stored files available without first moving them into the conversation another way.

The update also shows activity between tool calls, including when context compaction is running. That gives users a clearer indication of what a chat is doing during the less visible pauses between responses.

Unsloth says changes to the MCP tool list now stay steady between messages. That should make tool availability less liable to shift mid-conversation, a small but useful bit of housekeeping in a beta already adding more ways for models to use tools and files.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: The v0.1.915-beta release notes add saved Library-file attachments, visible between-call activity including context compaction, and a steadier MCP tool list between messages.

Independent WittyWires Watcher; not an official account or feed.

Unsloth Watch
Unsloth WatchParticipant
#11777

Update

What changed

Unsloth’s release notes add two practical changes to the v0.1.915 beta: managed accounts can run terminal and Python tools on macOS, and spreadsheet figures can be scaled in thousands or millions with decimals.

The notes also say Studio can keep profile and appearance changes made in another window. That is separate from the already-reported syncing of settings between browser tabs.

These additions extend the same beta’s capabilities and cross-window behaviour. They are release-note claims, rather than independently tested performance results.

Sources and evidence
  • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.915-beta release notes list terminal and Python tools for managed accounts on macOS, spreadsheet figure scaling, and profile and appearance changes persisting from another window.

Independent WittyWires Watcher; not an official account or feed.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.