AI Potluck
Contribute
Back to Gap Map Model components / Fine-tuning code

Unsloth

Unsloth AI
open core / Overall score: 3.5

Unsloth is a fine-tuning library that trains language, diffusion, speech and embedding models with custom Triton kernels for faster training and lower VRAM use. Supported methods include LoRA, QLoRA, full fine-tuning, pretraining, GRPO reinforcement learning and DPO preference tuning, and trained models export to formats including GGUF. Unsloth runs as code through Unsloth Core, or through the Unsloth Studio web UI and a desktop app.

Openness

4 high confidence
4.0
license
Apache-2.0(OSI, core unsloth pkg)+AGPL-3.0(OSI, optional Unsloth Studio UI)
source
public(unslothai/unsloth)
core features withheld
yes — unsloth.ai/pricing sells Unsloth Pro and Enterprise on capabilities the free open-source tier does not have: the free column lists MultiGPU as coming soon while Pro lists Enhanced MultiGPU support and up to 8 GPUs, and Enterprise adds multi-node, full training and 5x faster inference

Unsloth's core package is Apache-2.0 and the optional Studio UI is AGPL-3.0, both OSI licenses, with the full source public. But Pro and Enterprise tiers add capabilities the free package lacks -- multi-GPU support beyond one card, multi-node training and full rather than LoRA-only training -- which withholds part of the product itself rather than just hosted convenience. That keeps an otherwise open project at open core rather than open source.

  • https://github.com/unslothai/unsloth recorded 2026-08-11

    Live public repository (HTTP 200); GitHub's own page describes it as "The local UI to run and train text and diffusion models, including Kimi K3, Gemma 4, Qwen3.6, DeepSeek-V4, FLUX and more." and lists both an Apache-2.0 license and an AGPL-3.0 license for the repo.

  • https://raw.githubusercontent.com/unslothai/unsloth/main/LICENSE recorded 2026-08-11

    LICENSE body states: "Files under unsloth/*, tests/*, scripts/* are Apache 2.0 licensed. Files under studio/*, unsloth_cli/* which is optional to install are AGPLv3 licensed."

  • https://unsloth.ai/pricing recorded 2026-08-11

    Three tiers. Free (Open-Source) lists "Open-source", "Supports Mistral, Gemma", "Supports LLama 1, 2, 3", "MultiGPU - coming soon", "Supports 4 bit, 16 bit LoRA". Unsloth Pro lists "2.5x number of GPUs faster than FA2", "20% less memory than OSS", "Enhanced MultiGPU support", "Up to 8 GPUS support". Unsloth Enterprise lists "32x number of GPUs faster than FA2", "up to +30% accuracy", "5x faster inference", "Supports full training" and multi-node support. Both paid tiers are contact-for-pricing.

Adoption

3 high confidence
3.0

Adoption is measured by PyPI downloads of unsloth. GitHub stars corroborate continued use of the repository.

Capability

4 high confidence
4.0

Unsloth's custom Triton kernels make training faster and cut VRAM use, with the largest gains the project reports coming on mixture-of-experts models. Method coverage is broad, spanning full fine-tuning, LoRA, QLoRA, pretraining, GRPO reinforcement learning, DPO preference tuning and embedding fine-tuning, though its edge is memory and speed on one machine rather than the multi-node parallelism Megatron-LM offers.

  • https://raw.githubusercontent.com/unslothai/unsloth/main/README.md recorded 2026-09-28

    README states "Train LLMs, diffusion, TTS, and embedding models 2× faster with 70% less VRAM" and "Supports [reinforcement learning](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide), LoRA, QLoRA, full fine tuning, pretraining, RL, GRPO, DPO, and FP8." and "Train **MoE LLMs 12x faster** with 35% less VRAM", plus "Multi GPU setups" support.

Verified 2026-09-28, except some scores, still awaiting re-confirmation