Intel / repo

unsloth v0.1.915-beta — Improved Customization + Glass + Browser Use

Source: github.com
Author: shimmyshimmer
Date: Published Nov 4, 2026

Context

Most of Unsloth's v0.1.915-beta release changes its desktop app, but the change with the most technical weight is an opt-in FP8 mode for training. LoRA fine-tunes a model by training small adapter matrices while the original weights stay frozen. In the new mode, described in the FP8 LoRA pull request, the frozen base layers do the matrix multiplications for their forward pass and input gradients in 8-bit floating point, while the adapters keep their existing precision. It applies to 16-bit base models, not 4-bit ones, and is switched on with float8_training = "fp8_lora". The release notes aim it at recent NVIDIA GPUs, and the PR says T4, A100 and AMD ROCm setups are refused before the model loads.

The PR's author measured it on Qwen3-4B, with six fresh runs per setting and 2,048-token sequences. The median training step fell from 1.2218 to 0.9096 seconds, which the author reports as 1.3431 times faster, and peak GPU memory allocated during training fell from about 30,928 to 20,967 MiB, or 32.2% lower. The cost shows up in system memory: host RAM rose from about 9,153 to 13,423 MiB because exact 16-bit copies of the base weights are kept in system memory. On 80 held-out text windows, the difference in loss between the two modes had a confidence interval that includes zero. The author's wording is that "no accuracy difference is resolved", which means the test did not detect one, not that the two are proven equivalent. Windows, Apple GPUs and multi-GPU runs were not measured.

Other changes affect the app's memory and model handling. Unsloth no longer loads PyTorch and other machine-learning libraries at startup, which the notes say significantly reduces idle memory, and the startup features can be configured. Switching away from a GGUF model and back keeps the chat context, so a long history does not have to be re-read. Models in Apple's MLX format load and reload faster, and DPO, ORPO, KTO, GRPO and CPO preference and reinforcement training now run on MLX. With browser use enabled, models can click and type in HTML and React pages created in chat and edit files in the built-in browser, with a Back button that restores the previous version. The Glass themes introduced in the previous release are now labelled beta.

The release also adds Unsloth's Dynamic V3.1 versions of Qwen3.8-Flash-Next, which the linked guide describes as a 125-billion-parameter open-weight model from Qwen. The release highlights claim "10-15% extra accuracy" for these quants without naming a baseline. The guide measures how much of the full model's accuracy each quant retains and compares divergence against Unsloth's older V3 quants, but it does not state the 10-15% figure, so the sources do not establish what that number refers to.

Key takeaways · AI-distilled

FP8 LoRA training is switched on with float8_training = "fp8_lora", running an FP8 forward and input gradient against a frozen 16-bit base on recent NVIDIA GPUs.
Unsloth no longer imports PyTorch and other ML libraries at startup, cutting idle memory, and startup warmup is now opt-in per feature under Settings > General > Startup performance.
Switching away from a GGUF model and back preserves chat context, so a long history no longer has to be re-read on return.
MLX gains DPO, ORPO, KTO, GRPO and CPO training through trl, plus training from a saved adapter, continued , and streamed and audio datasets with exact resume.
With browser use enabled, models can click and type in HTML and React pages built in chat and read or edit HTML, Markdown, code and text files in the browser panel, where Back restores the previous version.

Terms in this piece · Glossary

LoRA — A cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.
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.
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.

Read the source github.com