Intel / repo

unsloth v0.1.916-beta — Large Performance Fixes

Source: github.com
Author: shimmyshimmer
Date: Published Nov 6, 2026

Context

The headline speedup in Unsloth's v0.1.916-beta release uses a second GPU to make image and video generation faster for models that already fit on one card. Unsloth is an open-source toolkit for and running models on your own hardware, and its app also runs image and video models. With two matching, idle NVIDIA GPUs, the release notes say SDXL, Qwen-Image, Wan2.2 5B and similar models run half of each generation step on the second card, about 1.5 to 1.8 times faster. The pull requests behind it call this CFG-parallel. Classifier-free guidance, or CFG, is how many image models steer toward a prompt: each denoising step runs the model once with the prompt and once without it, then combines the two predictions. The two passes do not depend on each other, so with a copy of the model on each card they can run at the same time. For FLUX and Lumina 2 the app instead splits an even-sized batch of images between the two cards, and the Lumina 2 pull request reports 1.7 to 2.1 times faster.

Much of the rest of the release removes slowness from the app itself: typing in a long chat no longer lags when a Glass theme or chat background is on, chat search keeps working on very long histories, and SDXL loads up to 40% faster. For people who use Unsloth as a Python library, the CUDA-graph text generation that v0.1.914 turned on in the app is now the default in model.generate() for Llama, Qwen2 and Qwen3 models on NVIDIA GPUs under Linux, which the notes say makes small models up to several times faster. The notes also claim Unsloth's new Dynamic V3.1 versions of Qwen3.8-Flash use 50% fewer and are 15% more accurate, without naming a baseline or pointing to a measurement.

Key takeaways · AI-distilled

With two matching, idle NVIDIA GPUs, Unsloth now runs half of each diffusion step on the second card for models such as SDXL, Qwen-Image and Wan2.2 5B, about 1.5 to 1.8x faster per the release. FLUX and Lumina 2 instead split an even-sized batch across both cards.
Each chat in Unsloth now gets its own built-in browser for the model to drive, so multiple chats can browse independently, and the browser can float over the chat as a movable, resizable window.
On 16 to 24 GB GPUs, LTX-2.5 now encodes new prompts on the GPU instead of the CPU, so clips with a new prompt render much faster. Its GGUF Q8_0 build also renders about 10% faster on NVIDIA GPUs with identical frames.
In the Python library, model.generate() for Llama, Qwen2 and Qwen3 models is faster by default on NVIDIA GPUs under Linux, up to several times faster for small models according to the release notes.
The release fixes a training bug in Assistant completions only mode: answers that began with a newline had been left out of training.

Terms in this piece · Glossary

fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
quantization — Shrinking a model by storing its numbers less precisely — like rounding — so it runs faster and fits on smaller hardware, at a small quality cost.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.

Read the source github.com