Intel / repo

unsloth v0.1.913-beta — Customize + Performance + UI/UX improvements

Source: github.com
Author: shimmyshimmer
Date: Published Oct 30, 2026

Context

The change most users of Unsloth's desktop app will notice in v0.1.913-beta is a reversed tradeoff in how updates install. The app is written partly in Python, and Python can compile its source files into bytecode either ahead of time or the first time they run. The previous release, v0.1.912-beta, advertised a first launch twice as fast. This release makes that install-time precompile opt-in, which the notes say cuts update time roughly in half. The cost is stated plainly: the first launch after an update is slower again, as it was before v0.1.912.

The other substantive change is for people models, which route each token through a few specialist sub-networks instead of the whole model. and QLoRA are cheap fine-tuning methods that train a small add-on, called an adapter, instead of rewriting the model's weights. Unsloth now runs several compatible LoRA or QLoRA jobs on the same mixture-of-experts model, such as Qwen3.6-35B-A3B, together through its fine-tuning API, with each job keeping its own adapter. The merged changes describe computing every job's expert adapters in one grouped pass and giving each batched job its own loss scaling and its own accounting. The notes do not give a throughput figure for this feature.

Most of the remaining speedups come from Unsloth's own tests on specific hardware: Linux chat scrolling rising from about 17 to 60 frames per second on NVIDIA with X11 and driver 610 or newer, Qwen-Image-2512 rendering up to twice as fast under a 16 GB VRAM limit, and four simultaneous Claude Code, Codex and OpenCode requests on a Mac finishing up to 2.3 times faster when they share an MLX batch. The new DeepSeek-V4.1-Flash-1017 files are labelled a preview, and the claim that they are the most accurate for their size on nearly all benchmarks is Unsloth's own, made in these release notes, which link to the model page. An experimental Neural Engine prefill option on Apple Silicon is opt-in, and Unsloth warns it can slightly change replies.

Key takeaways · AI-distilled

Unsloth v0.1.913-beta cuts app update time roughly in half by skipping install-time Python precompilation. The tradeoff is a longer first launch after updating, as before v0.1.912.
Keeping the Markdown parser linear after an unmatched $$ cut parsing of an 83 KB reply from 71 seconds to 0.7 seconds. Previewing a 1 GB local Parquet file now reads only its leading rows and peaks near 300 MB instead of 1.5 GB.
On Mac, Claude Code, Codex and OpenCode requests can share an MLX batch; in Unsloth's test four simultaneous requests finished up to 2.3x faster. An opt-in Neural Engine prefill ran prompts up to 31% faster on M3 Max but can slightly change replies.
Compatible LoRA and QLoRA jobs on MoE models such as Qwen3.6-35B-A3B can now train together through the fine-tuning API, each with its own adapter.
FLUX.2-klein-9B now runs on Linux hosts with 12.7 GB of system RAM by spilling its int8 denoiser to a scratch file, and Qwen-Image-2512 renders up to 2x faster under a 16 GB VRAM limit.

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.
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.
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.

Read the source github.com