# Unsloth v0.1.913-beta improves Linux performance and adds MoE fine-tuning

> **Key takeaway:** Unsloth's latest beta release optimizes Linux UI responsiveness, accelerates image generation under VRAM constraints, and introduces batched Mixture-of-Experts (MoE) fine-tuning.

**Published:** 2026-10-30T15:10:31+00:00 
**Source:** Unsloth Fast LLM Finetuning 
**Category:** ai-local-edge 
**Canonical URL:** https://fosswire.org/news/unsloth-v01913-beta-improves-linux-performance-and-adds-moe-.html 

## Summary
Unsloth has released version 0.1.913-beta, focusing on performance optimizations for Linux users and enhanced capabilities for local AI model training. The update significantly improves the user interface experience on Linux, specifically targeting smoother scrolling for long chats when using NVIDIA hardware with X11 and driver 610 or newer.

Technical improvements extend to model execution and training, including faster rendering for Qwen-Image models under 16 GB VRAM limits and the ability to run compatible LoRA and QLoRA jobs on Mixture-of-Experts (MoE) models simultaneously via the fine-tuning API. The release also introduces support for DeepSeek-V4.1-Flash-1017 via Unsloth Dynamic v3.1 GGUFs and enables FLUX.2-klein-9B to run on Linux systems with as little as 12.7 GB of system RAM.

## Technical improvements in Unsloth v0.1.913-beta
- **Linux Graphics Optimization:** Achieved a jump in high-resolution chat scrolling from approximately 17 fps to 60 fps on Linux systems utilizing NVIDIA drivers 610+ and the X11 windowing system.
- **VRAM-Constrained Inference:** Optimized Qwen-Image-2512 and Qwen-Image-Edit-2511 to render up to 2x and 1.3x faster, respectively, when operating under a 16 GB VRAM limit.
- **MoE Fine-Tuning:** Implemented batched fine-tuning for Mixture-of-Experts models, allowing compatible LoRA and QLoRA jobs to run together through the API, each maintaining its own adapter.
- **Memory and Parsing Efficiency:** Reduced peak memory usage for 1 GB Parquet file previews from 1.5 GB to 300 MB and optimized math formatting parsing to prevent long delays in long-form text streaming.
- **Apple Silicon Enhancements:** Introduced an experimental Neural Engine Prefill that can increase prompt processing speeds by up to 31% on M3 Max hardware.

## Why it matters
This release lowers the barrier for running high-parameter models like FLUX.2 on consumer-grade Linux hardware with limited RAM. By optimizing MoE fine-tuning and VRAM-constrained inference, Unsloth provides local AI developers with more efficient tools for training and deploying sophisticated models without requiring enterprise-grade GPU clusters.