builders · review · first seen Nov 7, 10:31 UTC

Unsloth adds per-chat browsers and performance fixes

Unsloth shipped separate browsers for chats, a movable browser window, and image and video performance changes. The reason to look now is practical: you can test whether these changes remove friction from your own chat and generation workflow.

L0 · primary1 source4 claims checked5 min read

What shipped

Separate chats can now run their own browsers. That is the relevant change if you keep multiple chats in progress and want a browser associated with each one. The release notes describe separate browsers for each chat; they do not establish that those browsers have separate security boundaries, stored data, or network settings. Treat this as a workflow feature, not a claim about isolation. [1]

Unsloth also shipped a built-in browser window that you can move and resize over a chat. This gives you a way to keep the browser visible while working in that chat, rather than treating the browser as a fixed part of the layout. The release notes say you can let the model work with the floating browser, but they do not specify which browser actions it can complete. Test the actions your workflow needs rather than assuming a broader set. [1]

For image work, Unsloth says SDXL loads up to 40% faster. The measured activity in that claim is loading, not image generation. “Up to” also matters: it does not promise the same reduction for every setup or every load. If loading interrupts your work, this is worth measuring locally. If your bottleneck is what happens after the model loads, this claim alone does not resolve it. [1]

Supported image and video models can use two GPUs under specific conditions: the NVIDIA cards must match, both must be idle, and the model must fit on one card. The release notes describe a potential speed benefit from two-GPU generation, but the supplied claim does not establish a result for a particular model or machine. Check those conditions before treating a second card as useful capacity for this release. [1]

Why it matters if you build

The browser changes are most useful to test with a task that crosses several chats. Open separate chats, give each one browser work, and see whether the arrangement makes it easier to keep the tasks straight. Then move and resize the floating window during an actual chat. The shipped controls make that trial possible today. They do not, by themselves, tell you whether the arrangement will suit a crowded screen or a long-running task. [1][2]

Keep the two performance questions separate. For SDXL, time the part of your workflow spent waiting for a load; the claim is about that interval. For image or video generation on two cards, first check that you have matching, idle NVIDIA GPUs and a supported model that fits on one. A faster load would not establish faster generation, and eligibility for two-GPU use would not establish the improvement on your workload. [1][2]

The costs and limits are still part of the decision. The supplied material does not quantify memory use for separate browsers, resource use for the floating window, or the cost of keeping two GPUs idle and available. It also does not name the supported image and video models in the supplied claim. If those details determine whether you can adopt the release, make the trial narrow and record your own constraints before changing a regular workflow. [1][2][3]

How to try it today

1. Run browser tasks in two separate chats and check how each browser follows its chat.
2. Move and resize the built-in browser over a chat while doing a real task.
3. If you use SDXL, compare its load time in your own workflow. Keep generation time separate.
4. If you have two matching NVIDIA GPUs, check that both are idle and that your supported image or video model fits on one card.
5. Try that model with both cards, then compare the result with your usual setup before changing your routine.

What we checked

Every claim below was walked down the source ladder. The rung is how close the source is to the thing itself; the tense is the source's own.

1. shipped L0 · primary

   Unsloth lets separate chats run their own browsers.

   per GitHub releases · unslothai/unsloth (L0)
2. shipped L0 · primary

   Unsloth adds a built-in browser window that can be moved and resized over a chat.

   per GitHub releases · unslothai/unsloth (L0)
3. shipped L0 · primary

   Unsloth says SDXL loads up to 40% faster.

   per GitHub releases · unslothai/unsloth (L0)
4. shipped L0 · primary

   With two matching, idle NVIDIA GPUs, supported image and video models can use both cards when they fit on one.

   per GitHub releases · unslothai/unsloth (L0)

What is not confirmed

The supplied claims do not establish browser isolation, persistence, or the full range of actions the model can perform in the floating window. Separate per-chat browsers and a movable window are confirmed; those additional properties are not. [1][2]

The material does not establish a typical SDXL load-time improvement, a generation-time improvement from the SDXL loading change, or a two-GPU result for your model. It does not provide a complete supported-model list or measurements of resource use. Those are questions for a local test, not conclusions to carry into a purchasing or deployment decision. [1][2]

Who should act, who should wait

Act now if your work regularly spans several chats, if you want a browser visible over a chat, or if SDXL loading is a noticeable interruption. These are shipped changes you can check against a real task without assuming that the reported performance applies to you. If you already have matching, idle NVIDIA GPUs and an eligible model, test two-card generation as a separate experiment. [1][2][3][4]

Wait before committing a workflow that depends on browser isolation, a particular model being supported on two GPUs, or a specific performance gain. The supplied material does not confirm those requirements. In particular, do not treat “up to 40% faster” SDXL loading as a general image-generation result, and do not assume that two GPUs help when the stated hardware, idle-state, or model-fit conditions are unmet. [1][2][3]

Sources

• Large Performance Fixes — GitHub releases · unslothai/unsloth L0 · primary