{"full_name":"unslothai/unsloth","owner":"unslothai","name":"unsloth","shard":"1","url":"https://github.com/unslothai/unsloth","description":"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.","homepage":"https://unsloth.ai/docs","topics":["fine-tuning","llama","llms","gemma","unsloth","llm","deepseek","text-to-speech","tts","qwen","agent","openai","reinforcement-learning","self-hosted","ui"],"tag_pages":["agent","deepseek","fine-tuning","gemma","llama","llm","llms","openai","qwen","reinforcement-learning","self-hosted","text-to-speech","tts","ui","unsloth"],"language":"Python","license":"Apache-2.0","avatar":"https://avatars.githubusercontent.com/u/150920049?v=4","created_at":"2023-11-29T16:50:09Z","pushed_at":"2026-11-16T09:16:52Z","metrics":{"stars":79591,"forks":7424,"watchers":378,"issues":5237,"prs":12148,"releases":87,"issues_open":159,"issues_closed":5078,"prs_open":273,"prs_closed":1685,"prs_merged":10190,"discussions":234},"growth":{"day":{"stars":94,"forks":12,"base_day":"2026-11-15"},"week":{"stars":363,"forks":46,"base_day":"2026-11-09"},"month":{"stars":1715,"forks":230,"base_day":"2026-10-17"}},"history":{"dates":["2026-09-19","2026-09-20","2026-09-21","2026-09-22","2026-09-23","2026-09-24","2026-09-25","2026-09-26","2026-09-27","2026-09-28","2026-09-29","2026-09-30","2026-10-01","2026-10-02","2026-10-03","2026-10-04","2026-10-05","2026-10-06","2026-10-07","2026-10-08","2026-10-09","2026-10-10","2026-10-11","2026-10-12","2026-10-13","2026-10-14","2026-10-15","2026-10-16","2026-10-17","2026-10-18","2026-10-19","2026-10-20","2026-10-21","2026-10-22","2026-10-23","2026-10-24","2026-10-25","2026-10-26","2026-10-27","2026-10-28","2026-10-29","2026-10-30","2026-10-31","2026-11-01","2026-11-02","2026-11-03","2026-11-04","2026-11-09","2026-11-10","2026-11-11","2026-11-12","2026-11-13","2026-11-14","2026-11-15","2026-11-16"],"stars":[75854,75946,75994,76028,76065,76158,76205,76239,76308,76377,76439,76537,76585,76639,76710,76773,76828,76929,77026,77119,77133,77169,77231,77271,77343,77506,77626,77707,77876,77940,78002,78064,78121,78184,78269,78322,78375,78425,78474,78521,78595,78631,78694,78756,78819,78963,79008,79228,79275,79324,79374,79392,79445,79497,79591],"forks":[6901,6910,6921,6928,6932,6941,6948,6946,6955,6970,6980,6996,7007,7016,7025,7041,7052,7070,7083,7093,7090,7097,7114,7119,7129,7148,7160,7173,7194,7197,7203,7209,7216,7224,7241,7245,7255,7263,7268,7283,7295,7301,7306,7316,7322,7354,7360,7378,7382,7384,7395,7398,7407,7412,7424],"watchers":[382,382,382,382,384,384,384,384,385,384,384,384,384,385,385,385,385,385,385,384,384,384,385,387,387,388,389,389,386,385,385,387,387,386,387,386,386,386,386,381,382,382,381,380,378,379,379,380,380,377,377,377,377,378,378],"issues_open":[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,160,160,160,167,159,159],"prs_open":[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,328,328,328,337,273,273]},"boards":{},"same_owner":[{"full_name":"unslothai/notebooks","name":"notebooks","shard":"1","boards":{}}],"issues":[{"number":2755,"title":"[FEATURE] TPU Support for Fine Tuning","comments":13,"created_at":"2025-06-18T00:53:47Z"},{"number":17402,"title":"Qwen3.8-Flash-Next UD 3.1 GGUFs vs current MTP heads: draft acceptance drops ~74% -> 56-66% at production temperature (Ryzen AI Max+ 395) — are 3.1-matched heads planned?","comments":6,"created_at":"2026-11-10T13:22:10Z"},{"number":17015,"title":"Studio: repeated full prompt processing across turns with Qwen hybrid model after tool use","comments":6,"created_at":"2026-11-07T11:57:02Z"},{"number":16317,"title":"[Bug] Unsloth North Mini Code giving Stream interrupted error when using Bearer Token and OpenAIAPI with OpenCode","comments":6,"created_at":"2026-10-31T17:20:46Z"},{"number":9037,"title":"[Bug] Unsloth Studio / Desktop: long chats redo the full prefill after a model reload because slot KV is only saved on idle auto-unload","comments":6,"created_at":"2026-08-16T19:04:05Z"},{"number":5764,"title":"[Feature Request] Multi GPU fine tuning via Unsloth Studio","comments":6,"created_at":"2026-05-25T10:11:22Z"},{"number":4380,"title":"[Feature] Flatpak support - Unsloth Studio","comments":6,"created_at":"2026-03-17T20:20:13Z"},{"number":2739,"title":"[Feature] Add ABBA: Highly Expressive Hadamard Product Adaptation for Large Language Models","comments":6,"created_at":"2025-06-12T14:30:20Z"}],"discussions":[{"number":17570,"title":"loraspectra: Fast Rust CLI to audit spectral rank, bake rsLoRA scaling, compress, and merge Unsloth LoRA .safetensors","comments":0,"category":"Show and tell","created_at":"2026-11-12T02:30:50Z"},{"number":17412,"title":"How do I add third-party decision (systemone) models?","comments":2,"category":"Q&A","created_at":"2026-11-10T14:35:11Z"},{"number":16817,"title":"Unsloth Desktop issues with using Qwen 3.8 27B with NVFP4","comments":4,"category":"Q&A","created_at":"2026-11-05T15:32:42Z"},{"number":16779,"title":"OpenUI support","comments":0,"category":"Ideas","created_at":"2026-11-05T11:50:08Z"},{"number":16487,"title":"When will Unsloth exit Beta state?","comments":0,"category":"Q&A","created_at":"2026-11-02T20:01:44Z"},{"number":16460,"title":"Run big llm-models like Qwen3.8-Flash-Next in q3 or iq4 like on strata/poystrata on 48GB+16GBVRam","comments":1,"category":"Q&A","created_at":"2026-11-02T06:54:05Z"},{"number":16456,"title":"load error","comments":2,"category":"Q&A","created_at":"2026-11-02T05:21:25Z"},{"number":15852,"title":"VRAM & KV-Cache Estimator for GGUF Export & Post-Fine-Tuning Serving (Zero-Dependency CLI & Interactive WebApp)","comments":0,"category":"Show and tell","created_at":"2026-10-28T20:15:02Z"}],"latest_release":{"name":"Dynamic v3.1 GGUFs + Performance Fixes","tag":"v0.1.917-beta","published_at":"2026-11-11T16:00:56Z","body":"Qwen3.8-27B is here with Dynamic v3.1 GGUFs, ready to run locally with Unsloth. This release also brings major performance improvements, smoother Glass rendering, revamped code blocks and Thinking UI, HTML/PDF chat exports, and new image generation and coding agent features.\r\n\r\n## Highlights\r\n- **New Qwen3.8-27B** Dynamic v3.1 GGUFs.\r\n- **Faster chats:** Major performance improvements to streaming, scrolling, menus and dialogs, especially in long conversations.\r\n- **Smoother Glass:** Reduced lag while typing, scrolling and generating.\r\n- **Revamped code blocks and Thinking:** Independently scrolling, expandable panels with improved syntax highlighting.\r\n- **HTML and PDF exports:** Export chats directly from the chat menu.\r\n- **Image queue:** Queue upcoming image generations while another is rendering.\r\n- **Claude Code and Codex on MLX:** Run coding agents with MLX models through `unsloth start`.\r\n\r\n## Chat & UI\r\n\r\n- Streaming performance no longer degrades as conversations grow.\r\n- Smoother scrolling with background images, with or without Glass.\r\n- Faster menus, dialogs and browser pane loading.\r\n- Read aloud starts after the first sentence instead of waiting for the entire response.\r\n- Chat preferences, time format and spell check now sync with your account.\r\n\r\n## Images & Video\r\n\r\n- FLUX.2 and FLUX.3 use **36–43% less VRAM** during image decoding at the same speed.\r\n- Queue image generations from the Images page or API.\r\n- INT8 and FP8 image models now load on older NVIDIA GPUs, including T4.\r\n- Fixed black image outputs with FLUX.3.\r\n- LTX-2.5 FP8 now runs on 24 GB GPUs.\r\n- Cosmos3-Nano can continue video clips from their first or last frames.\r\n\r\n## Coding Agents & API\r\n\r\n- Run Claude Code and Codex with MLX models on Mac using `unsloth start`.\r\n- New Unsloth MCP preset lets coding agents access Unsloth tools.\r\n- Claude API models without an explicit token limit no longer stop at 1,024 tokens.\r\n- MLX models now return a proper context-length error instead of an empty response.\r\n\r\n## Training\r\n\r\n- `model.generate()` is now **2–8x faster on T4 GPUs** by default across FastModel models.\r\n- Video columns are now correctly supported when training vision models.\r\n- Ulysses context parallelism now supports Kimi-Linear, DeepSeek, MiniMax and Gemma 4 text models.\r\n\r\n## What's Changed\r\n* Bump install.sh / install.ps1 pins to unsloth>=2026.11.3 by @danielhanchen in https://github.com/unslothai/unsloth/pull/16920\r\n* Studio GGUF: keep the stock Q8_0 forward on torch below 2.7 by @danielhanchen in https://github.com/unslothai/unsloth/pull/16911\r\n* Studio: merge LoRA and LoKR adapters into ComfyUI-quantized weights by @oobabooga in https://github.com/unslothai/unsloth/pull/16874\r\n* Studio: restore WebKitGTK offscreen code fences in batches and on any pointer release by @danielhanchen in https://github.com/unslothai/unsloth/pull/16919\r\n* Studio: warm the compile workers and Triton / inductor keys during a load that compiles inline by @danielhanchen in https://github.com/unslothai/unsloth/pull/16918\r\n* Studio: pre-seed the float specialisations of an image load's first compiled render by @danielhanchen in https://github.com/unslothai/unsloth/pull/16917\r\n* Studio diffusion: reuse the early quant probe's result instead of spawning a second child by @danielhanchen in https://github.com/unslothai/unsloth/pull/16910\r\n* Studio: keep Glass fast on WebKit by giving glass surface contents their own layers by @danielhanchen in https://github.com/unslothai/unsloth/pull/16921\r\n* Studio: release a streamed encoder's safetensors mapping once its weights are pinned by @danielhanchen in https://github.com/unslothai/unsloth/pull/16914\r\n* Studio: load long chat history faster on Linux by mounting parked history in larger steps by @danielhanchen in https://github.com/unslothai/unsloth/pull/16924\r\n* Studio: decode escaped spreadsheet text during RAG indexing by @DimitriChrysafis in https://github.com/unslothai/unsloth/pull/16712\r\n* Studio: refuse MLX prompts ","is_prerelease":false},"readme":true,"readme_zh":false,"news":[{"title":"Show HN: 80% faster, 50% less memory, 0% loss of accuracy Llama finetuning","url":"https://github.com/unslothai/unsloth","hn":"https://news.ycombinator.com/item?id=38487199","source":"Hacker News","lang":"en","date":"2023-12-01","points":385,"comments":119},{"title":"Unsloth: 80% faster, 50% less memory, 0% accuracy loss Llama finetuning","url":"https://github.com/unslothai/unsloth","hn":"https://news.ycombinator.com/item?id=38492652","source":"Hacker News","lang":"en","date":"2023-12-01","points":132,"comments":3},{"title":"Show HN: Finetune Llama-3.1 2x faster in a Colab","url":"https://colab.research.google.com/drive/1Ys44kVvmeZtnICzWz0xgpRnrIOjZAuxp?usp=sharing","hn":"https://news.ycombinator.com/item?id=41050468","source":"Hacker News","lang":"en","date":"2024-07-23","points":16,"comments":2},{"title":"Unsloth Desktop: Open-Source App for Local Models","url":"https://github.com/unslothai/unsloth","hn":"https://news.ycombinator.com/item?id=49262440","source":"Hacker News","lang":"en","date":"2026-08-11","points":10,"comments":2},{"title":"Show HN: Unsloth Studio - Local Fine-tuning, Chat UI","url":"https://github.com/unslothai/unsloth","hn":"https://news.ycombinator.com/item?id=47414372","source":"Hacker News","lang":"en","date":"2026-03-17","points":8,"comments":2},{"title":"Unsloth: Introducing AMD support","url":"https://github.com/unslothai/unsloth/releases/tag/v0.1.50-beta","hn":"https://news.ycombinator.com/item?id=48983714","source":"Hacker News","lang":"en","date":"2026-07-20","points":7,"comments":0},{"title":"Show HN: Open-source fine-tuning in a Colab notebook","url":"https://colab.research.google.com/drive/1lN6hPQveB_mHSnTOYifygFcrO8C1bxq4?usp=sharing","hn":"https://news.ycombinator.com/item?id=41305917","source":"Hacker News","lang":"en","date":"2024-08-21","points":5,"comments":0},{"title":"Finetune language models 30x faster","url":"https://unsloth.ai/introducing","hn":"https://news.ycombinator.com/item?id=38474791","source":"Hacker News","lang":"en","date":"2023-11-30","points":2,"comments":0},{"title":"What Happens When a Trusted Model Repo Changes? Unsloth Studio Re-Checks Before It Runs","url":"https://www.marktechpost.com/2026/10/07/what-happens-when-a-trusted-model-repo-changes-unsloth-studio-re-checks-before-it-runs/","source":"MarkTechPost","lang":"en","date":"2026-10-06"},{"title":"Unsloth Studio Flaw Turns Routine Model Inspection Into Code Execution","url":"https://www.darkreading.com/application-security/unsloth-studio-flaw-model-inspection-code-execution","source":"Dark Reading","lang":"en","date":"2026-09-29"},{"title":"通义千问携手Unsloth Dynamic V3，Qwen3.8-27B量化升级，低内存设备也能畅跑","url":"https://www.msn.cn/zh-cn/news/other/%E9%80%9A%E4%B9%89%E5%8D%83%E9%97%AE%E6%90%BA%E6%89%8Bunsloth-dynamic-v3-qwen38-27b%E9%87%8F%E5%8C%96%E5%8D%87%E7%BA%A7-%E4%BD%8E%E5%86%85%E5%AD%98%E8%AE%BE%E5%A4%87%E4%B9%9F%E8%83%BD%E7%95%85%E8%B7%91/ar-AA2aAn2L","source":"ITBear科技资讯 on MSN","lang":"zh","date":"2026-08-20"},{"title":"無料でローカルAIモデルのメモリ使用量を最大80％削減し処理速度を2倍に向上させる「Unsloth Studio」、Windows・macOS・Linux対応で ...","url":"https://gigazine.net/news/20260318-unsloth-studio/","source":"GIGAZINE","lang":"zh","date":"2026-03-18"},{"title":"Unsloth 推出 Browser use，Qwen3.8-27B 在本地操控浏览器","url":"https://news.google.com/rss/articles/CBMilAFBVV95cUxNN0pPMnA1VkpJaU01MkMzOFdFSE5jU1BqUzV5aXVZU0lsZzFNTWVaZ3RSZllkbkpDY0hBSGJ4bktpeE54NndhcWhfTE9SXzdDN19XTXBQVjNsOW1yOHNETEQ3b3dFcXR2RnlqazZDWk55UXA4ZFhpYk5Ld2pKYkpla1ZOOUdiSlhDSERzSUlFaElIc3Vj?oc=5","source":"Pasquale Pillitteri","lang":"zh","date":"2026-10-28"},{"title":"Unsloth Desktop：首款可本地运行训练模型应用，Dynamic 3.0 GGUFs 量化技术重大升级！-CS","url":"https://news.google.com/rss/articles/CBMia0FVX3lxTE56MWV1MFItWUJFMjBwOUhWcVNOOUhZckw0ZEstOUo1eWh4cGtuZHlEcjFqZ244YnVDenplQTBZQjRNSTQ2UDNlcVYwQnhtZ3VhUnZwT0xVQXRoN0RYMjFIYkdMSTRhNEVZOHFn?oc=5","source":"CSDN博客","lang":"zh","date":"2026-08-20"},{"title":"兄弟俩用开源对抗AI训练垄断，还顺手帮模型们修了一堆Bug","url":"https://news.google.com/rss/articles/CBMiUkFVX3lxTE9GZy1WUWVydzdhaXBlbDhSU2huaTNaenprVDdhUWJsMnhlOEE2OEhOUVhYZURoczE4UlRXdVR6UzRZdEdqV3BscWJycWM3RE9CTHc?oc=5","source":"麻省理工科技评论","lang":"zh","date":"2025-10-27"},{"title":"从RLHF、PPO到GRPO再训练推理模型，这是你需要的强化学习入门指南","url":"https://news.google.com/rss/articles/CBMif0FVX3lxTE5XMmFsQ3JBWTNzbXRidWZ1WUdBajlLS0FhVDN4SHlxYWpnOERNbUk1SlhoT0o3TFZETU56NTJtWVVqdmZ1NzROMjF0VUdEVVRqcVkyLVIxQ195UXFvcWJYTERTTFQwakdOVXp0MmxkaElXOGxFQkNMeFhqNE5IVXM?oc=5","source":"新浪财经","lang":"zh","date":"2025-06-22"},{"title":"社区供稿 | Unsloth 微调 Llama3-8B，提速 44.35%，节省 42.58% 显存，最少仅需 7.75GB","url":"https://news.google.com/rss/articles/CBMiSEFVX3lxTFBoSld6S0ZFYU9pb1ozaDlHNk81RDVZRnZvWlNINnZETEZ6ZUNCSGFkVjlLc1plYURKRHBMNXo1U3VUNEFIbmdZVw?oc=5","source":"智源社区","lang":"zh","date":"2024-05-09"}],"stale":false,"snapshot_at":"2026-11-16T09:23:14Z","history_start":"2026-09-06"}