Which models does Unsloth support for fine-tuning?

Unsloth supports fine-tuning models including Llama, Mistral, Gemma, Qwen, and GLM. It is designed to work with the Hugging Face TRL and PEFT libraries.

How much faster is Unsloth compared to standard training methods?

Unsloth claims to train models 2-30x faster and use up to 90% less VRAM. The editor take notes that real, measurable 2-5x speedups are common on consumer GPUs.

Is Unsloth free to use?

The core library is free and open-source, covering 4-bit and 16-bit LoRA on a single GPU. Paid Pro and Enterprise tiers, which offer multi-GPU and multi-node support, require contacting sales.

Can I export models trained with Unsloth to other formats?

Yes, Unsloth allows for clean export to GGUF, vLLM, and Ollama. A newer no-code desktop app also supports local training and inference for Mac and Windows.

Is Unsloth a managed cloud service?

No, Unsloth is a training accelerator, not a managed cloud service. Users must provide their own GPU and MLOps infrastructure, as it does not offer hosted inference endpoints or a full UI.

Unsloth

✓ Editorially verified

Open-source LLM fine-tuning toolkit with custom kernels that train 2-30x faster and use up to 90% less VRAM.

Freemium· Free open-source; Pro and Enterprise contact salesFine-tuningLlama, Mistral, Gemma, Qwen, GLM (multi-model)8.2 / 10

Visit website →

In short

Unsloth is a fine-tuning framework that accelerates LoRA and QLoRA training on single consumer GPUs using custom Triton kernels. It is best for developers seeking the fastest, most memory-efficient path to fine-tune open-weights models like Llama and Mistral within the Hugging Face ecosystem.

Best for

Pick Unsloth if you're fine-tuning open-weights LLMs on your own GPU and want the fastest, most memory-efficient path that still works with the Hugging Face stack.

Skip if

Skip it if you want a fully managed cloud fine-tuning service with hosted inference and a UI, or if you only ever use closed-model APIs.

Unsloth is a fine-tuning framework built around custom Triton kernels and manually rewritten autograd math, letting you LoRA/QLoRA-train models like Llama, Mistral, Gemma, Qwen and GLM on a single consumer GPU. The free open-source library plugs into Hugging Face TRL and PEFT and is the default 'fast path' in dozens of Colab notebooks for instruction-tuning, DPO/ORPO, vision and audio adapters. A newer no-code desktop app adds local training and inference for Mac and Windows, with dataset generation from PDFs/CSVs and export to GGUF, vLLM and Ollama.

The open-source tier is genuinely usable on its own and covers 4-bit and 16-bit LoRA on a single GPU; the paid Pro and Enterprise tiers unlock multi-GPU (up to 8), multi-node, higher-accuracy kernels and faster inference. Pricing for those isn't published — it's contact-sales — which signals the commercial focus is on labs and companies running serious training jobs rather than hobbyists. Hobbyists will be fine on the free version forever.

Unsloth has become something of a default in the open-weights fine-tuning community because the speedups are real and the notebooks 'just work'. The main caveat is that it's a training accelerator first — if you want a managed cloud fine-tuning service with a UI and hosted inference endpoints, this isn't that.

Editor's take

Unsloth is one of the few open-source fine-tuning projects whose performance claims hold up under scrutiny — the kernels are the real deal. For anyone training Llama, Mistral, Gemma or Qwen on a single 24GB card, it's effectively the default choice. The paid tiers are aimed at labs, not weekend hackers, and that's fine.

— The AI Tool Bible editorial team

Pros

✅ Real, measurable 2-5x speedups and big VRAM savings on consumer GPUs
✅ Open-source core with permissive license and active GitHub
✅ Drop-in compatible with Hugging Face TRL, PEFT and transformers
✅ Excellent ready-to-run Colab notebooks for most popular models
✅ Exports cleanly to GGUF/llama.cpp, vLLM and Ollama

Cons

⚠️ Multi-GPU and multi-node are gated behind paid tiers with opaque pricing
⚠️ Not a hosted service — you still bring your own GPU and MLOps
⚠️ Cutting-edge model support sometimes lags official releases by days

Use cases

lora-finetuningqloralocal-trainingdpo-orpomodel-quantizationgguf-export

Frequently asked

Which models does Unsloth support for fine-tuning?: Unsloth supports fine-tuning models including Llama, Mistral, Gemma, Qwen, and GLM. It is designed to work with the Hugging Face TRL and PEFT libraries.
How much faster is Unsloth compared to standard training methods?: Unsloth claims to train models 2-30x faster and use up to 90% less VRAM. The editor take notes that real, measurable 2-5x speedups are common on consumer GPUs.
Is Unsloth free to use?: The core library is free and open-source, covering 4-bit and 16-bit LoRA on a single GPU. Paid Pro and Enterprise tiers, which offer multi-GPU and multi-node support, require contacting sales.
Can I export models trained with Unsloth to other formats?: Yes, Unsloth allows for clean export to GGUF, vLLM, and Ollama. A newer no-code desktop app also supports local training and inference for Mac and Windows.
Is Unsloth a managed cloud service?: No, Unsloth is a training accelerator, not a managed cloud service. Users must provide their own GPU and MLOps infrastructure, as it does not offer hosted inference endpoints or a full UI.