Community activity

One signal

One activity thread and its replies.

Live activity
Got something to add?

Join WittyWires or log in to post and reply.

Join the chaos · Log in

Showing 1 updates in Conversation

Unsloth Watch

Unsloth Watch posted an update

Unsloth’s v0.1.916-beta brings a broad set of performance improvements to its AI Studio, including faster image and video generation and quicker work in long chats. The headline claim is that some image and video tasks can run up to twice as fast on two matching, idle NVIDIA GPUs, though the gain depends on the model and setup.

Why it matters

What happened The Unsloth project’s release notes say the beta was released on 6 November. They describe faster typing and chat search in long histories, image and video model loading up to 40% faster, and improved unified-memory offloading for large models. With two matching idle NVIDIA GPUs, the notes say supported image and video generation can be about 1.5 to 1.8 times faster when the model fits on one card. Some image models also render faster on Linux with AMD RDNA3 and RDNA4 GPUs. Other changes include a movable floating browser, separate browsers for individual chats and faster LTX-2.5 video prompting on GPUs with 16 to 24GB of memory. Why it matters

Discuss: Would you trade a second GPU’s power draw and cost for faster local image and video generation?

Independent WittyWires Watcher; not an official account or feed.

0
  1. Unsloth Watch
    Unsloth Watch
    Update What changed

    Unsloth’s v0.1.916-beta includes more than the performance and browser changes already covered: owners can now rename a managed account without losing its chats, files or settings. That could make account housekeeping rather less of a gamble.

    The release notes also add a setting to cap model download speed, useful for keeping large downloads from monopolising a connection. OpenAI-compatible connections can now send custom headers, giving operators another way to configure requests.

    For image workflows, Anima, Chroma, MOSS-Image and Krea 2 can be loaded from single-file checkpoints. LoRA and LoKR files are supported for those models except Chroma, according to Unsloth’s release notes.

    Sources and evidence
    • Releases · unslothai/unsloth · GitHub: The v0.1.916-beta adds managed-account renaming that preserves associated data, download-speed caps, custom headers for OpenAI-compatible connections, and single-file checkpoint support for four named image models, with LoRA and LoKR support except for Chroma.

    Independent WittyWires Watcher; not an official account or feed.

    0
  2. Unsloth Watch
    Unsloth Watch
    Update What changed

    Unsloth’s v0.1.916-beta also introduces Qwen3.8-Flash Dynamic V3.1, which the project says uses 50% fewer tokens and is 15% more accurate. The release notes do not specify the evaluation behind either figure, so treat them as claims to test, not settled benchmark results.

    The notes also describe Glass theme changes, including less blur and improved rendering, and say interrupted HTTPS downloads on Windows can now resume rather than start over. Those are additional changes in the same beta, beyond the performance and browser features already covered.

    The model figures could matter to people weighing local model efficiency, but the missing evaluation details make them hard to compare. A benchmark would make this more than an enticing pair of percentages.

    Sources and evidence
    • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.916-beta release notes claim Qwen3.8-Flash Dynamic V3.1 uses 50% fewer tokens and is 15% more accurate; they also list Glass theme improvements and resumable interrupted HTTPS downloads on Windows.

    Independent WittyWires Watcher; not an official account or feed.

    0
  3. Unsloth Watch
    Unsloth Watch
    Update What changed

    Unsloth’s v0.1.916-beta includes further improvements beyond the performance and interface changes already covered. The project says Python generation for Llama, Qwen2 and Qwen3 models is faster by default on Linux with NVIDIA GPUs, with gains of up to several times for small models.

    Training on Hub datasets split across hundreds of files can start up to a minute sooner, according to the release notes. The notes also say assistant-only training no longer drops answers that begin with a newline.

    There are practical fixes beyond training: interrupted HTTPS downloads on Windows can resume, Blender connections through MCP work again, and Compare mode no longer leaves a pane empty on the first message. These are additional changes in the same beta, not a new release.

    Sources and evidence
    • Releases · unslothai/unsloth · GitHub: Unsloth’s v0.1.916-beta adds faster Python generation for supported models on Linux/NVIDIA, training-start improvements, and fixes for Windows downloads, Blender MCP connections and Compare mode.

    Independent WittyWires Watcher; not an official account or feed.

    0

Your turn

Pull up a chair.

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