{"id":1499,"date":"2026-09-18T00:15:57","date_gmt":"2026-09-17T15:15:57","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/09\/18\/unsloth-v0-1-810-beta-released\/"},"modified":"2026-09-18T21:42:06","modified_gmt":"2026-09-18T12:42:06","slug":"unsloth-v0-1-810-beta-released","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/18\/unsloth-v0-1-810-beta-released\/","title":{"rendered":"Unsloth v0.1.810-beta Released with Multi-User and AMD Support"},"content":{"rendered":"

<\/p>\n

At a Glance<\/h2>\n
\n\n\n\n\n\n\n\n\n\n
Item<\/th>\nValue<\/th>\n<\/tr>\n<\/thead>\n
Repository<\/td>\nunslothai\/unsloth<\/a><\/td>\n<\/tr>\n
Version<\/td>\nv0.1.810-beta<\/a><\/td>\n<\/tr>\n
Published<\/td>\n2026-09-17<\/td>\n<\/tr>\n
License<\/td>\nApache-2.0<\/td>\n<\/tr>\n
Source type<\/td>\nPrimary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n

Values determined by this site’s code at collection time. Dates are JST.<\/em><\/p>\n

<\/p>\n

Overview<\/h2>\n

The latest version “v0.1.810-beta” of unslothai\/unsloth, an open-source tool for running and fine-tuning diffusion models such as LLMs (Large Language Models) and image generation in local environments, has been released.<\/p>\n

unslothai\/unsloth is a Python-based project published under the Apache-2.0 license. It has gained 76,294 stars on GitHub and is popular as a local UI and fine-tuning tool supporting a wide range of models and formats, including GGUF, MLX, Qwen3.8, DeepSeek-V4, MiniMax-H3, Gemma 4, and FLUX.<\/p>\n

This update brings crucial feature additions and improvements for local LLM users, including multi-user account support, a revamped Docker image, expanded hardware support such as AMD RDNA1\/RDNA2 and Windows ARM64, and enhanced chat and reasoning capabilities.<\/p>\n

Key Changes<\/h2>\n

Introduction of Multi-User Accounts<\/h3>\n

A multi-user account feature has been added to enable usage by multiple people in local environments or shared servers. Accounts can be created from “Settings > Accounts” using a one-time setup code and individual passwords. Each user’s chats, projects, credentials, and training data are isolated separately, but loaded models can be shared and used efficiently if settings match. The existing single-user behavior is maintained as is.<\/p>\n

Revamped Docker Image<\/h3>\n

The Docker image has been updated to reduce image size, and Studio data is now persisted on volumes. This makes it easier to update, generated passwords, configure custom ports, and retain packages without locking down application code. Additionally, Unsloth’s own training patches have been restored on GPU hosts, and locally running LM Studio, Ollama, and Hermes models can be automatically detected from the Docker installation environment. Furthermore, an AMD ROCm image for Linux hosts has been newly added alongside the CUDA version.<\/p>\n

Enhanced Chat and Reasoning Capabilities<\/h3>\n

In the chat UI, queued prompts can now be edited, reordered, and steered. This is operable even while a local model is loading. Additionally, reasoning budgets settings for GGUF models, desktop UI interface scaling, chat width adjustments, and a more compact composer have been added. Responsiveness during long-running reasoning executions, chat history saving, and handling of files and images generated by tool execution have also been improved.<\/p>\n

Enhanced Security and Sandboxing<\/h3>\n

Credential protection and tool call validation have been strengthened. Approval requests will now appear when accessing files outside the sandbox. Furthermore, OS-level tool sandboxing has been added as a preview for supported Linux and macOS hosts. Network access is not restricted, but the automatic mode maintains existing protections even when OS-level isolation is not available.<\/p>\n

Training and Installation Improvements<\/h3>\n

A dataset download feature from “Data Recipes” has been added, and bugs related to 16-bit full fine-tune exports, LoRA saving, and Hugging Face’s “Push to Hub” have been fixed. Checkpoint saving during container shutdown, memory estimation during training, and dataset processing have also been improved. In the core section, fixes were made to normalization, RoPE, Q-GaLore, and distributed device selection. Installation reliability improvements on Windows environments, GPU package repairs, and offline installation verification were also implemented.<\/p>\n

Research and API Feature Updates<\/h3>\n

In the “Deep Research” feature, unfinished reports are now saved more appropriately, and JSON output processing and reports for executions where no evidence was obtained have been improved. When uploading scanned PDFs, fallback to local OCR functions. Video input support at the \/v1\/chat\/completions<\/code> endpoint and access to installed Ollama models via API have also been added.<\/p>\n

Supported Models and Hardware<\/h2>\n

This version adds and improves support related to the following hardware and models:<\/p>\n