Unsloth v0.1.916-beta Released: Performance & Multi-GPU Updates

At a Glance
| Item | Value |
|---|---|
| Repository | unslothai/unsloth |
| Version | v0.1.916-beta |
| Published | 2026-11-06 |
| License | Apache-2.0 |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
A new version of unslothai/unsloth, v0.1.916-beta, has been released. unslothai/unsloth is a Python-based local UI and tool (License: Apache-2.0) for running and training LLMs and diffusion models.
This release brings numerous updates centered around performance, including improved display and search performance in long chats, faster loading and rendering for image and video models, and parallel generation using two GPUs.
In particular, this update significantly improves usability and generation speed for users running image and video generation on setups with two identical NVIDIA GPUs, as well as those working with long chat histories.
Key Changes
Faster Image and Video Generation Using Two GPUs
In environments where two identical NVIDIA GPUs are installed and idle, running models such as SDXL, Qwen-Image, and Wan2.2 5B will now split half of each step across the second card even if the model fits on a single card. This speeds up processing by approximately 1.5x to 1.8x.
Additionally, for FLUX and Lumina 2, even-sized image batches are split and executed across the two cards. This directly impacts users who have multi-GPU environments and generate content using diffusion models. This feature is automatically enabled on qualifying setups.
Improved Loading and Output Speeds for Image and Video Generation Models
Loading times for SDXL models are up to 40% faster, reducing the wait time before the first image is generated.
When using 16 to 24 GB GPUs, LTX-2.5 models now encode new prompts on the GPU instead of the CPU, speeding up clip display for new prompts. Furthermore, rendering LTX-2.5 GGUF Q8_0 videos on NVIDIA GPUs is about 10% faster while maintaining output frame quality.
Support has also been added for loading image models such as Anima, Chroma, MOSS-Image, and Krea 2 from single-file checkpoints.
Improvements to Long Chats and UI Performance
In long chat histories, typing lag no longer occurs even when the Glass theme or chat backgrounds are enabled. Furthermore, chat search and deletion operations complete quickly without failing, even when very long conversation histories are present.
An issue where streaming responses would freeze for a few seconds the first time audio features were used after startup has also been resolved. This improves workflow efficiency for users who accumulate long-term conversations and large volumes of messages in the chat feature.
UI Feature Expansion and Floating Browser
A feature has been added allowing the built-in browser to float as a small, resizable window over the chat screen, with adjustable positioning and size. This floating state is preserved per chat.
Additionally, the display of code blocks and thinking blocks has been revised and placed inside scrollable containers. Optimization of the Glass theme—replacing background image blurring with a darkening process—delivers up to a 2x response speed improvement on certain Linux desktop environments that do not use GPU rendering.
Python Inference Optimizations and Bug Fixes
When using unsloth as a Python library on Linux with NVIDIA GPUs to run model.generate() for Llama, Qwen2, and Qwen3 models, decoding with CUDA graphs is now enabled by default, improving generation speed primarily for smaller models.
Other fixes include ensuring unified memory offloading functions properly with large models, resolving out-of-memory errors with Auto context settings when running GGUF models on CPU-only environments, and fixing update-related bugs on Windows.
Supported Models and Hardware
This version adds support for new image and video generation model architectures, single-file checkpoints, and quantization formats, significantly expanding local inference and generation options.
Newly Supported Models and Checkpoint Formats
- Anima / Chroma / MOSS-Image / Krea 2: Added support for loading ComfyUI-format single-file checkpoints (.safetensors). Except for Chroma, these models also support applying LoRA and LoKR files. Krea 2 also supports loading int4 quantized checkpoints.
- Qwen-Image-2.1-Turbo: Added support for the NVFP4 quantization format, implementing quantization processing based on dedicated layer policies.
- Other Added Model Features: Introduced speedups for the MiniMax-H3 video decoder, faster recovery times for SAM 3 via memory pinning, and the maintenance of thinking processes in the gemma-4-thinking template.
Quantization Formats and Accelerator Support
- NVFP4 / FP8 / int8 (W8A8): Added support for loading NVFP4.safetensors formats in diffusion models, alongside compiled int8 W8A8 linear calculation kernels for ROCm.
- NVIDIA GPU: Supports parallel execution using SDXL, Qwen-Image, Wan2.2 5B, FLUX, and Lumina 2 in environments with two identical idle GPUs. CUDA graph-based decoding is also enabled by default, boosting inference speeds for Llama, Qwen2, and Qwen3.
- AMD RDNA3 / RDNA4 GPU (ROCm): Added support for installing flash-attn via the Triton AMD backend on Linux environments. FLUX and Z-Image rendering are accelerated, and Qwen-Image-2.1 achieves execution speeds surpassing bf16 during int8 W8A8 operations.
- 16–24 GB VRAM Environments: New LTX-2.5 prompt encoding processing can now run on the GPU instead of the CPU, shortening the time until video clip rendering begins.
- CPU Inference Environments: Auto context memory settings for GGUF models have been optimized, and logical CPU thread allocation improvements have boosted processing speeds by 1.03x to 1.26x when running via sd.cpp or PyTorch on the CPU.
How to Get It
Unsloth is available as a desktop app (“Unsloth Desktop"), a Web UI (“Unsloth Studio"), and a codebase (“Unsloth Core").
Updating Unsloth Desktop
Pre-built installers are available for Windows, macOS, and Linux (Ubuntu deb / AppImage). Get the latest binary for your OS from the official releases page or download link and install it.
Update Procedure for Unsloth Studio (Development Version / Source Code)
If you build and use it from existing source code, update the Git repository and then run the installation script.
macOS / Linux / WSL (Bash environment):
cd unsloth && git pull
./install.sh --local
unsloth studio -p 8888
Windows (PowerShell environment):
cd unsloth; git pull
.\install.ps1 --local
unsloth studio -p 8888
If you isolate your build in a specific directory (UNSLOTH_STUDIO_HOME), update by supplying the environment variable.
macOS / Linux / WSL:
cd unsloth && git pull
UNSLOTH_STUDIO_HOME="$PWD/.studio"./install.sh --local
UNSLOTH_STUDIO_HOME="$PWD/.studio" unsloth studio -p 8888
Windows:
cd unsloth; git pull
$env:UNSLOTH_STUDIO_HOME="$PWD\.studio";.\install.ps1 --local
$env:UNSLOTH_STUDIO_HOME="$PWD\.studio"; unsloth studio -p 8888
If you want to pre-specify the inference backend for GGUF chat execution (vulkan, cuda, rocm, cpu, etc.), set environment variables such as UNSLOTH_LLAMA_CPP_BACKEND=vulkan before running the script.
Related Articles
- Unsloth v0.1.917-beta Released with Major Performance Boosts
- Unsloth v0.1.907-beta Released with Major Speed and UI Upgrades
- Unsloth v0.1.811-beta Released with AMD & NVIDIA Docker Images
- Unsloth v0.1.918-beta Adds Computer Use, MCP Game Engine Support
What to Read Next
- Follow this tool → Unsloth overview and release history (89 releases tracked)
- Other quantization, model formats and fine-tuning → ggml / ExLlamaV3
Sources
- unslothai/unsloth v0.1.916-beta Release Notes
- PR #16682: Studio diffusion: Automatic partner card, FLUX batch split, LoRA scale mirror, CFG batching fallback
- PR #16691: Decode with CUDA graphs by default in plain generate(), with memory release and an eager fallback
- PR #16710: Studio: load Anima, Chroma, MOSS-Image and int4 Krea 2 ComfyUI single files
- PR #16767: Studio: install flash-attn’s Triton AMD backend on ROCm RDNA3 / RDNA4
- PR #16825: Studio diffusion: compiled int8 W8A8 linear on RDNA3 / RDNA4 ROCm
- PR #16775: Studio: the in-app browser can float over the chat as a small, resizable window

