◆ AI Model & Data Infrastructure

Unsloth

v0.1.804-beta open-source

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, Kimi K3, MiniMax-H3, Gemma 4, FLUX and more.

source ↗ website ↗ rss ↗ last repo push Aug 31, 2026

unsloth start claude --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL

Summary

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, Kimi K3, MiniMax-H3, Gemma 4, FLUX and more.

Release history

v0.1.804-beta Aug 27, 2026 · issue 012

Unsloth v0.1.804-beta adds Qwen3.8-Flash-Next and GLM-5.3-Flash local inference, JSONL chat export, sampling seed, and smarter GPU+RAM offloading.
└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.804-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.804-beta
```
- ›Adds a sampling seed setting in chat run settings for reproducible inference.
- ›Supports Qwen3.8-Flash-Next (125B) locally on 75 GB RAM via 1-bit Unsloth Dynamic GGUF, with up to 262K context and switchable reasoning levels (None, Low, Medium, Extra High).
- ›Supports GLM-5.3-Flash (320B, 18B active) locally on 102 GB combined RAM + VRAM via 1-bit GGUF, with up to 1M context and Low/High/Max reasoning modes.
- ›Enables 5x faster inference for RAM offloading, with large GGUFs automatically split across GPU and system RAM via the offload planner (now on by default).
- ›Shows estimated memory usage before loading a model in the Load Model panel.
+8 moreshow less
- ›Displays VRAM usage directly from the downloaded models list.
- ›Resumes local chats after a disconnect instead of losing the in-progress reply.
- ›Keeps Deep Research running when a provider applies rate limiting.
- ›Images returned by MCP tools now appear inline in chat.
- ›Supports searching and downloading embedding models directly from Hugging Face.
- ›Loads text-to-speech models on demand rather than at startup.
- ›Model settings are preserved when switching between chats.
- ›Supports 'Preserved Thinking' for Qwen3.8-Flash-Next to keep reasoning consistent across longer conversations.
v0.1.803-beta Aug 25, 2026 · issue 012

Unsloth v0.1.803-beta adds auto-compaction for long chats, LAN/remote access, custom llama.cpp builds, Intel XPU support, and OpenCode V2 via unsloth start.
└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.803-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.803-beta
```
└──▷ HOW TO FIND IT
Enable LAN access so teammates on the same network can reach your local model without Cloudflare.
📍Settings › Remote & LAN Access › Enable LAN access, then change the generated admin password and optionally enable auto-start.
- ›Adds unsloth start opencode to connect OpenCode V2 agents to local models.
- ›Adds --reasoning-effort flag documented on the Agents tab for controlling reasoning in Studio.
- ›Adds edit_file tool for partial file modifications in chat/agent workflows.
- ›New Auto Compaction (Experimental) feature lets long local chats continue past a model's context limit by rolling older turns out of active context, indexing evicted turns into a per-thread searchable archive via search_conversation without summarization.
- ›New Remote & LAN Access (Preview) settings section lets users enable or disable LAN access without restarting, with connection addresses, QR codes, and optional auto-start (disabled by default, requires changing the generated admin password).
+7 moreshow less
- ›Adds support for custom llama.cpp builds with toggles for Cache RAM, Mmap, Mlock, Checkpoints, Speculative Decoding KV Cache, and Vision On/Off.
- ›Adds Intel XPU support for inference and hardware acceleration.
- ›Responses API now supports structured output formatting.
- ›Adds custom STT (speech-to-text) connections in Studio.
- ›Adds drag-and-drop support for files (including .ods) onto project sources and knowledge bases.
- ›Hub now supports more image and video workflows.
- ›Releases Unsloth Dynamic v3.0 GGUFs for Qwen3.8-27B, delivering >10% higher top-1 accuracy.
v0.1.802-beta Aug 25, 2026 · issue 012

└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.802-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.802-beta
```
- ›Adds --reasoning-effort flag for controlling reasoning on the Agents tab.
- ›Adds search_conversation retrieval for evicted chat turns in the auto-compaction RAG pipeline.
- ›New Auto Compaction (Experimental) feature lets long local chats continue past the model context limit by rolling older turns into a per-thread searchable archive using the existing RAG pipeline (store, chunker, embedder, retrieval) — no summarization.
- ›New Remote & LAN Access (Preview) with a dedicated Settings section; LAN access can be enabled or disabled without restarting and supports QR codes and optional auto-start. Disabled by default and requires changing the generated admin password.
- ›Adds Intel XPU support.
+4 moreshow less
- ›Adds prompt queueing and customizable keyboard shortcuts.
- ›Projects feature lets chats, files, and workspaces be organized together; chats remember composer settings, system prompts, and model sampling options.
- ›Drag-and-drop support for .ods files and attachment preview modal for all attachment types.
- ›Keyless/password-less LAN API access now supported.
v0.1.801-beta Aug 20, 2026 · issue 012

└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.801-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.801-beta
```
└──▷ HOW TO FIND IT
Enable LAN access so teammates on the same network can reach your local model instance — no Cloudflare tunnel needed.
📍In the Unsloth Desktop or Studio UI, go to Settings › Remote Access and enable LAN access, then share the displayed connection address or QR code.
- ›Adds OpenCode V2 support for unsloth start opencode, extending the agent integration table.
- ›New Remote & LAN Access (Preview): enable or disable LAN access from a dedicated Settings section without restarting, with support for connection addresses, QR codes, and optional auto-start (disabled by default, requires changing the generated admin password).
- ›Adds Intel XPU support alongside existing ROCm, xFormers, and flash-attention backends.
- ›Adds prompt queueing with Cmd/Ctrl+Enter and drag-to-reorder, plus customizable keyboard shortcuts.
- ›Projects now organize chats, files, and workspaces together, with attached files shared across a project.
+1 moreshow less
- ›Chats now persist composer settings, system prompts, and model sampling options per model.
v0.1.800-beta Aug 14, 2026 · issue 012

Unsloth v0.1.800-beta adds Qwen3.8 support, custom llama-server args, FP8 MiniMax-H3, Codex tool calling, and live debug logs
└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.800-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.800-beta
```
└──▷ TRY IT
Run Qwen3.8-27B locally via Claude Code agent, pointing it at the GGUF quantization that fits in ~17 GB RAM.
```
$ unsloth start claude --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL
```
- ›Adds support for running and fine-tuning Qwen3.8-27B and Qwen3.8-2.4T locally, including Dynamic GGUF quantizations runnable in 17 GB RAM and NVFP4 quants.
- ›Model settings now expose an 'Extra Arguments' box for custom llama-server flags, validated against the installed build and saved per model for normal, startup, and API load modes — invalid or security-breaking flags are rejected with a clear message.
- ›Adds Fast FP8 MiniMax-H3 inference, delivering approximately 10x faster throughput (approximately 3 minutes vs. 30).
- ›External connected AI providers now support tool calling and tool results passed back to the model for multi-step tasks.
- ›Sign-in with a Codex subscription enables Codex tools inside Chat.
+14 moreshow less
- ›Adds Settings > Debugging panel showing main server and model-runner logs with live, 3-second, and manual refresh modes; passwords, tokens, and API keys are redacted in the viewer.
- ›Chat now shows live prompt and generation speeds during streaming.
- ›Paste a YouTube link in Chat to attach its transcript, including title, channel, duration, link, and caption language.
- ›Save a full chat or individual reply into a project's sources, preserving reasoning, tool calls, and citations.
- ›Inference is up to 10% faster with lower VRAM use and a tunable memory limit in supported configurations.
- ›Idle image and video models can optionally unload to free VRAM for Chat or Training.
- ›MiniMax-H3 now splits large model parts to fit smaller supported GPUs; the model picker hides H3 options the current hardware cannot run.
- ›RAG document indexing now uses the CPU by default, preventing large GPU memory blocks from remaining reserved.
- ›GGUF exports now check available disk space before starting a long merge, showing how much space is needed.
- ›On Kaggle, large saves can use the bigger temporary drive when the working folder is too small.
- ›Training now checks whether a model requires a newer Transformers version and prompts before installing it, and warns when a model must use 16-bit instead of 4-bit.
- ›Adds improved support for AMD RDNA 3, RDNA 4, and Strix Halo systems with safer multi-GPU ROCm device matching.
- ›Pinned models can be reordered by dragging them in the model list.
- ›Chat settings now persist with the conversation across remote sessions.
v0.1.702-beta Aug 13, 2026 · issue 012

Unsloth Desktop adds tool calling with self-healing, Deep Research, OpenAI-compatible API, and model export to NVFP4/GGUF
└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.702-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.702-beta
```
- ›Adds tool calling and web search support for external cloud providers (OpenAI, Anthropic) inside Unsloth Desktop.
- ›Adds Deep Research mode that autonomously plans and produces cited reports from unlimited private web search.
- ›Adds RAG support for bringing user-supplied files into local model context.
- ›Adds MCP (Model Context Protocol) tool integration for workflows that need external apps, data, or actions.
- ›Exposes an OpenAI-compatible API for serving local models to agents and other clients.
+7 moreshow less
- ›Adds support for connecting OpenAI and Anthropic as cloud model providers inside the Desktop app.
- ›Supports running Muse Glimmer 30B, Kimi K3, Qwen3.8, DeepSeek-V4 Flash 0731, and Gemma 4 locally.
- ›Adds MiniMax-H3 for local video generation, alongside image and video diffusion model support with up to 2x faster inference on supported workflows.
- ›Adds model export to NVFP4, GGUF, and other formats after no-code training.
- ›Adds Remote Access via Cloudflare HTTPS, publishing the Desktop app and its local APIs to any device.
- ›Adds multi-GPU support spanning NVIDIA, AMD, Intel, and Mac hardware, including AMD RDNA3/4 and Strix Halo.
- ›VRAM usage is now tunable from the UI.
v0.1.701-beta Aug 11, 2026 · issue 012

Unsloth Desktop launches as a cross-platform app to run, train, and serve AI models locally with an OpenAI-compatible API.
└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.701-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.701-beta
```
- ›Adds unsloth start claude --model <model>, unsloth start codex, unsloth start hermes, unsloth start openclaw, and unsloth start opencode subcommands to connect Claude Code, OpenAI Codex, and other agents to local models.
- ›Adds unsloth studio --secure to expose Unsloth via a free Cloudflare HTTPS link for global remote access.
- ›Adds unsloth studio -H 0.0.0.0 -p <port> flags for binding the Studio server to a custom host and port for LAN access.
- ›Adds unsloth studio reset-password subcommand for resetting the Studio password.
- ›Supports UNSLOTH_STUDIO_PASSWORD environment variable for headless, password-pre-set Studio starts.
+8 moreshow less
- ›Adds --disable-tools flag to unsloth studio to block server-side tool execution when exposing the instance publicly.
- ›Launches Unsloth Desktop — a native GUI app for Windows, macOS, and Linux — for no-code model running, training, and export.
- ›Adds an OpenAI-compatible local API for serving local models to agents and third-party clients, with optional OpenAI and Anthropic cloud provider connections.
- ›Supports self-healing tool calling that repairs malformed calls, plus sandboxed Python and Bash code execution for models.
- ›Adds Deep Research mode for planning and generating cited reports, alongside unlimited private web search and RAG over user-supplied files.
- ›Supports export of trained models to NVFP4 and GGUF formats in addition to FP8 and others.
- ›Adds video generation support via MiniMax-H3 locally, plus image and video diffusion model inference at up to 2× faster speed on supported workflows.
- ›Supports running models on CPU, multi-GPU, NVIDIA, AMD, Intel, and Apple Silicon (Mac) hardware.
v0.1.70-beta Aug 11, 2026 · issue 012

└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.70-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.70-beta
```
- ›Adds unsloth start claude --model <model>, unsloth start codex, unsloth start hermes, unsloth start openclaw, and unsloth start opencode subcommands (Unsloth Start) to wire local models into Claude Code, OpenAI Codex, and other agents with a single command.
- ›Introduces Unsloth Desktop — a native app for Windows, macOS, and Linux — to run, train, and deploy LLM, diffusion, TTS, and embedding models locally without writing code.
- ›Serves local models through an OpenAI-compatible API, enabling drop-in use with agents and any OpenAI-compatible client.
- ›Adds self-healing tool calling that repairs malformed calls and executes Python and Bash inside sandboxed environments so models can test code and verify output.
- ›Supports export of trained models to NVFP4, GGUF, and FP8 formats.
+4 moreshow less
- ›Supports MCP tool connections for workflows that require external apps, data, or actions.
- ›Adds cloud provider connections for OpenAI and Anthropic models alongside local models.
- ›Supports multi-GPU inference and training across NVIDIA, AMD, Intel, and Mac hardware, plus CPU-only operation.
- ›Adds LAN access to local models via Settings > API keys > LAN access.
v0.1.61-beta Aug 10, 2026 · issue 012

Unsloth v0.1.61-beta adds Muse Glimmer 30B support, MiniMax-H3 video generation, and preliminary image diffusion
└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.61-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.61-beta
```
- ›Adds support for Meta's Muse Glimmer 30B dense model — runnable and fine-tunable locally on 20 GB RAM/VRAM (Mac and GPU), under the Apache 2.0 license.
- ›Adds MiniMax-H3 video generation support.
- ›Adds preliminary image diffusion support.
- ›Chat now provides downloadable artifacts from generated responses.
- ›Revamped training page in Unsloth Studio.
v0.1.60-beta Aug 10, 2026 · issue 012

Unsloth v0.1.60-beta adds Muse Glimmer 30B support, Qwen3-ASR dictation, and independent parallel-chat prompt queues.
└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.60-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.60-beta
```
- ›Adds support for running and fine-tuning Meta's Muse Glimmer 30B on 20GB RAM/VRAM (Mac and GPU), available via Unsloth Dynamic quants under the Apache 2.0 license.
- ›Adds Qwen3-ASR dictation to Unsloth Studio, with cancellable model downloads and a download prompt before starting.
- ›Supports independent prompt queues across parallel chat sessions in Studio, enabling concurrent conversations without queue cross-contamination.
- ›Studio image and video pickers now populate from a curated catalog rather than raw model listings.
- ›Desktop update controls moved to Settings > General, consolidating update management in one place.
+2 moreshow less
- ›Studio now detects stalled Deep Research output streams and surfaces the condition to the user.
- ›Desktop warns before quitting during an in-progress update and skips the installer fallback when the user cancels.
v0.1.527-beta Aug 9, 2026 · issue 012

Unsloth v0.1.527-beta adds Qwen3-ASR dictation, independent parallel chat queues, and cancellable model downloads in Studio.
└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.527-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.527-beta
```
- ›Supports independent prompt queues across parallel chats, so concurrent conversations no longer share a single queue.
- ›Serves gated diffusion base models from ungated Unsloth mirrors, removing the need for user tokens to access those models.
- ›Populates the image and video model pickers from a curated catalog.
- ›Links the Train page across to image training.
- ›Moves Desktop update controls to Settings > General.
+6 moreshow less
- ›Enables Desktop draft publishing without a full rebuild.
- ›Bounds dataset_num_proc by available memory and treats a value of 1 as genuine single-process mode rather than no multiprocessing.
- ›Fetches sentence-transformers model subfolder weights directly instead of pruning them.
- ›Answers GGUF variant requests from the specific named cache snapshot and skips unreadable entries.
- ›Settles deferred compile-mode switches between training steps rather than mid-step.
- ›Reports the real cause when a model save or GGUF export fails, rather than a generic error.
v0.1.526-beta Aug 4, 2026 · issue 012

Unsloth v0.1.526-beta adds DeepSeek V4 Flash and Kimi K3 Dynamic GGUF support with multi-GPU layer offload and 2x faster DSpark inference.
└──▷ GET THIS VERSION
```
$ git clone --branch v0.1.526-beta https://github.com/unslothai/unsloth.git
# already have the repo? check out this version:
$ git checkout v0.1.526-beta
```
- ›Adds DeepSeek V4 Flash 0731 Dynamic GGUF support with quantization tiers UD-IQ1_S (83 GB), UD-IQ3_XXS (104 GB), UD-Q4_K_XL (155 GB), and UD-Q8_K_XL (162 GB) for lossless inference.
- ›Adds Kimi K3 Dynamic GGUF support for Moonshot AI's 2.8T-parameter MoE model (104B active parameters, native vision, 1M context window) with tiers UD-IQ1_S (595 GB), UD-Q4_K_XL (1.51 TB), and UD-Q8_K_XL (1.56 TB) for lossless inference.
- ›Enables DSpark inference by default for DeepSeek V4 Flash, delivering 2x faster inference.
- ›Automatically detects multi-GPU setups and offloads model layers to system memory for both DeepSeek V4 Flash and Kimi K3.
- ›Supports low, high, and max reasoning effort levels when running Kimi K3 (thinking-only model).
+7 moreshow less
- ›Adds stop and send buttons on the Studio chat dictation bar.
- ›Studio now applies saved per-model settings on API loads and adds an API monitor.
- ›Studio reports active training status in the UI.
- ›Desktop adds Check for Updates and Update Now options to Settings.
- ›Desktop shows the app version in Settings > About.
- ›Desktop supports dropping documents into a chat (non-GGUF files accepted).
- ›Auto-falls back to HTTP download if XET is stuck, improving reliability on Colab, low-memory, and high-memory/CPU systems.