[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$f2v0ah5c9xtc5y":3},{"code":4,"message":5,"data":6,"timestamp":75},10000,"success",{"full_name":7,"position":8,"description":9,"language":10,"stars":11,"forks":12,"stars_gained":13,"age_days":14,"hn_threads":15,"one_liner":43,"summary_md":44,"topic":45,"github_topics":48,"license":69,"homepage":70,"updated_at":71,"history":72},"unslothai\u002Funsloth",8,"Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.","Python",78557,7291,48,1061,[16,22,27,33,37],{"id":17,"title":18,"points":19,"comments":20,"posted_at":21},38487199,"Show HN: 80% faster, 50% less memory, 0% loss of accuracy Llama finetuning",385,119,"2023-12-01",{"id":23,"title":24,"points":25,"comments":26,"posted_at":21},38492652,"Unsloth: 80% faster, 50% less memory, 0% accuracy loss Llama finetuning",132,3,{"id":28,"title":29,"points":30,"comments":31,"posted_at":32},49262440,"Unsloth Desktop: Open-Source App for Local Models",10,2,"2026-08-11",{"id":34,"title":35,"points":8,"comments":31,"posted_at":36},47414372,"Show HN: Unsloth Studio - Local Fine-tuning, Chat UI","2026-03-17",{"id":38,"title":39,"points":40,"comments":41,"posted_at":42},48983714,"Unsloth: Introducing AMD support",7,0,"2026-07-20","A desktop app to run and fine-tune LLMs and diffusion models locally, aimed at people who want training without a cloud GPU.","Unsloth is best understood as a fine-tuning stack that now ships a friendly desktop face. If you already have a local GPU and want to adapt a model on your own data, it removes the usual notebook-and-dependency assembly work. If you mainly want to chat with models, the training machinery is more than you need.\n\n## The problem\n\nA developer with a local GPU wants a model that follows their style, speaks their domain vocabulary, or handles a specific extraction task. The manual route means installing CUDA-matched PyTorch, pulling a fine-tuning library, writing a training loop, managing quantization, and figuring out how to convert and serve the result. Each step is a place to get stuck, and most of the work has little to do with the data or the goal.\n\nUnsloth changes this by putting model download, training and inference behind one interface. The user selects a model, points at a dataset, starts training, and can then run the resulting model through the same app. The deliverable is a working local model plus the option to export it in a format that other tools can load.\n\n## What's different\n\nThe core claim is efficiency. The project originally gained attention with a Show HN titled around 80% faster training and 50% less memory with no accuracy loss for Llama fine-tuning. That claim matters most on constrained hardware: less memory can mean a larger model or batch fits on a single consumer GPU, and shorter runs make iteration affordable.\n\nA skeptical commenter in that original thread noted that the MLP and attention paths still call the same PyTorch operations as Hugging Face, and questioned where the speedup comes from [skeptical HN comment](https:\u002F\u002Fnews.ycombinator.com\u002Fitem?id=38494191:). A reply pointed to a custom autograd and memory optimizations and suggested the advantage might not last [reply on HN](https:\u002F\u002Fnews.ycombinator.com\u002Fitem?id=38495137:), while another linked the project's own explanation [project explanation](https:\u002F\u002Funsloth.ai\u002Fintroducing). The important point for a prospective user is that the efficiency claim is substantially the author's own, backed by reproducible benchmarks rather than an independent measurement supplied here. Treat it as a reason to measure on your hardware, not as a guarantee.\n\nThe second difference is scope. This is not only a training library: it ships a desktop app, a web UI, and the code-level core, and covers inference, tool calling, diffusion and video generation alongside training [README](https:\u002F\u002Fgithub.com\u002Funslothai\u002Funsloth\u002Fblob\u002Fmain\u002FREADME.md). That breadth is convenient, but it also means the project is a moving surface with optional components under a different license. The core stays Apache-2.0 while the Studio UI is AGPL-3.0 [licensing note](https:\u002F\u002Fgithub.com\u002Funslothai\u002Funsloth?tab=AGPL-3.0-2-ov-file), which matters if you plan to redistribute or embed the UI.\n\n## Who should try it\n\nGood fit: someone with an NVIDIA, AMD, or Intel GPU, or even a CPU-only setup, who wants to fine-tune an open model on their own data and then serve it locally without stitching together separate tools. The desktop installer path is the low-friction entry; the web UI under Unsloth Studio is the alternative if you would rather access it from a browser or a remote machine [Studio docs](https:\u002F\u002Funsloth.ai\u002Fdocs\u002Fnew\u002Fstudio\u002F). macOS users and CPU-only users are documented as supported, so a discrete GPU is not a hard prerequisite, though training expectations should scale with the hardware.\n\nPoor fit: teams that want a managed training service with SLAs, or users whose only goal is a polished chat client. The training and model-management surface will be overhead, and the desktop-first workflow does not replace a production inference platform.\n\n## Requirements and limitations\n\nThe main constraint is hardware. CPU-only and macOS support exist, but the efficiency story is aimed at local accelerators; without one, expect the training workflow to be slow. Multi-GPU and NVIDIA, AMD and Intel GPUs are documented [features section](https:\u002F\u002Fgithub.com\u002Funslothai\u002Funsloth\u002Fblob\u002Fmain\u002FREADME.md), but the review should check the exact backend for their card against the install docs rather than assume it.\n\nThe second constraint is install and upkeep. There are three install paths: the desktop app, the Studio web UI, and the code-level core [install docs](https:\u002F\u002Funsloth.ai\u002Fdocs). The developer and nightly installs build from `main`, which means you follow the project's update cycle and may hit unpolished code; uninstall scripts exist [uninstall docs](https:\u002F\u002Funsloth.ai\u002Fdocs\u002Fnew\u002Fstudio\u002Finstall#uninstall). Budget time for driver and CUDA-compatible PyTorch issues, which the project cannot fully eliminate. A third consideration is the license split: if you plan to ship a modified UI, check the AGPL obligations before building on it.\n\n## HotRepo's take\n\nHotRepo has only one observed daily-board day for this repository: 2026-10-25, when it showed 78,536 stars and a +48 daily gain. That is a single snapshot, not a trend, and it cannot show whether interest is rising or fading. The 2023 Show HN thread drew 385 points and 119 comments [Show HN thread](https:\u002F\u002Fnews.ycombinator.com\u002Fitem?id=38487199), but that reflects discussion volume on an older release, not current adoption.\n\nA sensible judgement: the project is mature in age and scope, and the efficiency claim is the thing to check on your own hardware before committing a training pipeline to it. Verify the current documented support matrix for your GPU and OS, and whether Studio's AGPL-3.0 license is acceptable for how you intend to use the UI.",{"slug":46,"name":47},"machine-learning","Machine Learning",[49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68],"agent","ai","chatgpt","deepseek","fine-tuning","gemma","image-generation","llama","llm","llms","openai","python","qwen","reinforcement-learning","self-hosted","stable-diffusion","text-to-speech","tts","ui","unsloth","Apache-2.0","https:\u002F\u002Funsloth.ai\u002Fdocs","2026-10-25T11:03:22Z",[73],{"date":74,"stars":11,"stars_gained":13},"2026-10-25",1792968167]