The archive · Developer & Business Tools · Technical decision · 2023-2026
Unsloth bets hand-written GPU kernels make fine-tuning a desktop app; 75k stars
Sydney brothers rewrote LLM training's hot paths in Triton to fine-tune 2x faster with 70% less VRAM - free and open, then YC-backed.
Unsloth
What the business is
Unsloth makes open-source software to run and train open LLMs and diffusion models locally - a core library (Apache-2.0), the Unsloth Studio web UI, and Unsloth Desktop - advertising 2x faster training with up to 70% less VRAM and no accuracy loss.
How it started
Australian brothers Daniel and Michael Han started Unsloth in late 2023 as an open-source side project. Daniel, a former Nvidia engineer who had built low-memory numerical libraries, rewrote the heaviest transformer training paths as hand-derived Triton kernels; their first notebooks showed a 7B LLaMA-style model fine-tuning on a free T4 GPU in a fraction of the time and memory of stock Hugging Face configurations.
What happened
Unsloth AI joined Y Combinator's Summer 2024 batch, with public records placing its seed round near $500K (AI Wiki). The project compounded through 2024-25 as Daniel published detailed bug reports on flagship open models - eight issues in Google's Gemma, tokenization defects in Llama 3, and a sliding-window-attention bug in Phi-3 - with fixes flowing back into transformers and llama.cpp. Star-history records 74 GitHub Trending days, best rank #1 on 2026-02-10 and Python #1 again on 2026-08-17. By 2026 the project had grown from Colab notebooks to a web UI and a native desktop app, and by mid-2026 passed ~67k stars with 10M+ monthly downloads of prequantized weights.
No ending yet — it is still running.
Background
Unsloth is an open-source library that fine-tunes large language models up to 2x faster while using up to 70% less GPU memory, with no accuracy loss. Its bet was technical: instead of waiting for better models, rewrite the hot operators of transformer training - attention, rotary position embeddings, normalization, and loss - as hand-derived backward passes in OpenAI Triton, fusing operations that PyTorch's autograd would otherwise run separately.
Australian brothers Daniel and Michael Han started the project in late 2023, and it spread through Colab notebooks that showed a 7-billion-parameter LLaMA-style model fine-tuning on a free T4 GPU. Unsloth AI joined Y Combinator's Summer 2024 batch, and the founders gained unusual visibility by publishing detailed bug reports on flagship open models - eight issues in Gemma, tokenization defects in Llama 3, and a sliding-window-attention bug in Phi-3 - with fixes adopted back into transformers and llama.cpp.
The visibility loop repeated whenever a major open model shipped: star-history records 74 days on GitHub Trending with a best rank of #1 on 2026-02-10 and Python #1 again on 2026-08-17. By 2026 the project had expanded from notebooks into Unsloth Studio's web UI and a native desktop app for running and training models, passed ~67k GitHub stars, and reported 10M+ monthly downloads of its prequantized weights on the Hugging Face Hub.
What has to be true
- The bottleneck was real and shared: fine-tuning and RL training of open models required multi-GPU servers, while most developers and students had one consumer card or a free Colab session.
- Hand-derived backward passes were a durable edge: each fused Triton kernel cut memory traffic that generic autograd could not, and the math stayed equivalent to a standard fine-tuning run.
- Publishing bug fixes for Gemma, Llama 3, and Phi-3 turned a small team into a trusted fixture of the open-weights ecosystem, so every new model release brought new users.
- Keeping the core free under Apache-2.0 while selling Studio and desktop conveniences let Unsloth monetize without breaking the open loop that generated its stars.
What can be applied
Win on constants, not models: make the free path dramatically faster and you become the default stack - every flagship open-model release re-ran Unsloth's benchmark onto GitHub's front page.
Aftermath
As of September 2026, Unsloth is live and expanding: the repository holds 75.4k stars at global rank #234, and the project now describes itself as the first desktop app to run and train models, supporting NVIDIA, AMD, Intel, CPU, and Vulkan backends plus connections to coding agents like Claude Code and Codex. The core stays Apache-2.0 while the Studio UI is AGPL-3.0, and downloads of prequantized weights on Hugging Face keep growing. Its known weakness was single-GPU focus versus multi-node frameworks like Axolotl, which the December 2025 release began addressing.
Sources
- unslothai/unsloth - 75.4k Stars - Global Rank #234
- GitHub - unslothai/unsloth: Local UI to run and train LLMs and diffusion models
- Unsloth | AI Wiki
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card