← All reportsPublic report

unslothai/unsloth

Local UI to run and train LLMs and diffusion models, including Qwen3.8, Kimi K3, MiniMax-H3, Gemma 4, DeepSeek-V4, FLUX and more.

75k stars Python View on GitHubprofiled 11d ago
Try GitZoid
400PRs this week
34Contributors
0Deps scanned
0Issues found
01 · Repo overview

How unsloth is put together

Unsloth is a monorepo combining a Python LLM fine-tuning library ('Unsloth Core': monkey-patched HuggingFace transformers/peft/trl models plus custom Triton-style kernels), a Typer-based CLI ('unsloth_cli'), and 'Unsloth Studio' — a local web/desktop product consisting of a FastAPI backend (studio/backend), a React 19 + Vite + Tailwind frontend (studio/frontend), and a Tauri 2 desktop shell (studio/src-tauri). Data flow: users drive training/export/inference through the CLI or the Studio UI, which calls FastAPI routes that delegate to the core library running GPU work in worker processes; an opt-in MCP server (fastmcp) wraps the same route functions for agent access behind a bearer token. Distribution spans PyPI (pip/uv), native installers (install.sh/install.ps1), Docker (unsloth/unsloth image), and packaged desktop binaries built via the Tauri release workflow.

Languages

PythonTypeScriptRustShellPowerShell

Frameworks

FastAPIUvicornPydantic v2Typer/ClickfastmcpReact 19Vite 8 (rolldown/oxc)Tailwind CSS 4TanStack Router/TableTauri 2Zustand

Datastores

SQLite (sqlite-vec for embeddings/RAG)Dexie (IndexedDB, frontend-local)filesystem model/checkpoint storage

Infrastructure

Docker (unsloth/unsloth image with Jupyter + SSH ports)Cloudflare tunnel for remote HTTPS accessLAN binding mode (-H 0.0.0.0)GitHub Releases for desktop binariesNVIDIA CUDA / AMD ROCm / Vulkan GPU runtimespre-commit CI

Major components

unsloth core models

Patches and loads HuggingFace model families (llama, qwen2/qwen3, gemma, mistral, diffusion, vision, RL/DPO) for fast LoRA/full fine-tuning.

unsloth.kernels

Hand-written fused GPU kernels (rms_layernorm, rope_embedding, swiglu, cross_entropy_loss, fp8, flex_attention) used by the patched models.

unsloth_cli

Typer console app ('unsloth') exposing training config loading, inference, model catalog, subagent/MCP integration for Claude Code/Codex, and Studio launch gating.

studio/backend

FastAPI/Uvicorn server providing training, export, inference, model, and data-recipe routes, hardware detection, Cloudflare tunnel/LAN access, and platform compatibility shims.

studio/backend MCP server

Opt-in FastMCP server wrapping existing route functions (status, training start/stop, checkpoint load, GGUF export) guarded by a constant-time bearer-token ASGI middleware.

studio/frontend

React 19 + Vite SPA (chat UI, model picker, data designer, settings) communicating with the backend over /api and /v1 proxies.

studio/src-tauri

Tauri 2 desktop shell packaging the frontend into Windows/macOS/Linux apps with updater, deep-link, and window-state plugins.

installers & setup scripts

install.sh/install.ps1 and studio/install_*.py bootstrap Python stacks, prebuilt llama/sd/whisper/node binaries, and pinned dependencies across platforms.

Over the past three weeks the Unsloth team shipped a large wave of improvements to Studio (their desktop app), focusing on chat reliability, model loading and memory handling across NVIDIA and AMD hardware, plus a new performance benchmarking suite called studiobench. The middle week was by far the busiest, adding features like inline web-search images, LoRA finetune defaults, and a real-path A/B benchmark simulator. The most recent week polished installer pins, keyboard shortcuts, secure local API access, and a long list of small crash-and-display fixes.

Week by week

2026-08-24A broad polish week for Studio: 87 commits covering installer updates, secure local API access, AMD GPU reporting fixes, and many small chat and download bug fixes.latest6 changes

Feature

Secure keyless local API access

Studio can now serve its API securely on localhost and private networks without needing an API key.

Feature

Keyboard shortcut registry filled out

The app's full set of keyboard shortcuts is now registered so they work consistently throughout Studio.

Fix

Accurate AMD VRAM reporting on Windows

AMD graphics memory is now read correctly on Windows by matching adapter counters to the right device.

Feature

Automatic embedding offload and multi-GPU split

Studio now decides how to spread models across GPUs and offload embeddings without asking the user.

Fix

Installer no longer rolls back successful installs

If the installer cannot write to a shell profile file, it no longer undoes an otherwise completed installation.

Chore

Updated install scripts and model mappings

Install scripts were bumped to newer package versions and Qwen3.8-27B was mapped to the official Unsloth repositories.

2026-08-17The busiest week with 298 commits: studiobench, a new real-path benchmarking tool, landed alongside chat improvements like inline web-search images, default-on LoRA finetune targets, and many context-window and memory fixes.6 changes

Feature

studiobench performance benchmarking

A new tool measures Studio's real-world speed and can compare two configurations side by side, with verdict tools that prove the numbers are genuine.

Feature

Inline web search images in chat

Images found through web search now appear directly inside the chat conversation instead of as separate links.

Feature

LoRA finetune targets enabled by default

Fine-tuning now automatically selects sensible target layers without requiring manual configuration.

Fix

Clearer context window errors

Studio now tells you exactly which part of your prompt does not fit in the model's context window.

Refactor

Kaggle T4 kernels packed into one

Two operations were merged into a single kernel so Studio can use the freed-up session for other work.

Fix

Auto context selection improved

The automatic context-length chooser got smarter and now remembers your per-model settings between sessions.

2026-08-10A focused week of 15 commits centered on chat queueing, training data tokenization, and a sweep of test and CI speedups.6 changes

Feature

Prompt queueing in chat

You can now line up multiple prompts with Cmd/Ctrl+Enter and drag them into a different order before they run.

Refactor

Faster transformers model-type sweep

Checking every supported model type now runs in parallel across CPU cores instead of one at a time.

Feature

Online dataset tokenization

Plain-text training runs now convert text to tokens on the fly rather than preparing it all upfront.

Fix

Fixed MTP performance under partial GPU offload

Embedded multi-token prediction no longer slows down when only part of the model fits on the GPU.

Fix

xFormers compatibility with flash-attn 4

Keeping xFormers working when flash-attn 4 is installed, while guarding against an integer overflow on long inputs.

Chore

Faster tests and cheaper CI

Repository tests were cut roughly in half and three CI workflows stopped paying for results they threw away.

03 · Security check

Dependencies and code review

0 dependencies scanned

Dependency advisories

Security Watch

No known advisories across 0 scanned dependencies.

No known advisories in the scanned dependencies.

Code review

No concrete code-level issues confirmed.

Get this report every week for your repos.

GitZoid learns each repo, reports what changed, and flags what needs attention. One flat price for the whole team.

$19 a month, flat · First 10 outputs free · No card required