Unsloth AI

We made a complete Guide on Reinforcement Learning (RL) for LLMs!

Learn about:
• RL's goal & why it's key to building intelligent AI agents
• Why o3, Claude 4 & R1 use RL
• GRPO, RLHF, DPO, reward functions
• Training your own local R1 model with Unsloth

🔗 docs.unsloth.ai/basics/reinf...

You can now run DeepSeek-R1-0528 locally with our Dynamic 1-bit GGUFs! 🐋

We shrank the full 715GB model to just 168GB (-80% size).

We achieve optimal accuracy by selectively quantizing layers.

DeepSeek-R1-0528-Qwen3-8B is also supported.

GGUFs: huggingface.co/unsloth/Deep...

Introducing Unsloth Dynamic v2.0 GGUFs!

v2.0 sets new benchmarks on 5-shot MMLU + KL Divergence. So, you can now run quantized LLMs with minimal accuracy loss.

For benchmarks, we built an evaluation framework to match official MMLU scores of Llama 4 & Gemma 3

Blog: docs.unsloth.ai/basics/dynam...

You can now run Llama 4 on your local device! 🦙

We shrank Maverick (402B) from 400GB to 122GB (-70%).

Our Dynamic 1.78-bit iMatrix GGUFs ensures optimal accuracy & size by selectively quantizing layers.

Scout + Maverick GGUFs: huggingface.co/collections/...
Guide: docs.unsloth.ai/basics/tutor...

The 1.58-bit quant fits in 131GB VRAM (2× H100s) for fast throughput inference at ~140 tokens/s.

For best results, use 2.51-bit Dynamic quant & at least 160GB+ combined VRAM + RAM.

Basic 1-bit & 2-bit quantization causes the model to produce repetition and poor code. Our dynamic quants solve this.

How to Run Deepseek-V3-0324 Locally DeepSeek's V3-0324 model is the most powerful open-source model rivalling GPT 4.5 and Claude 3.7. Learn to run the model with Unsloth Dynamic quants.

You can now Run DeepSeek-V3-0324 locally using our 1.58 & 2.51-bit Dynamic GGUF!🐋

We shrank the 720GB model to 131GB (-80%) by selectively quantizing layers for the best performance. Fixes basic quants breaking & bad output issues: www.unsloth.ai/blog/deepsee...

GGUF huggingface.co/unsloth/Deep...

We teamed up with 🤗Hugging Face to release a free notebook for fine-tuning Gemma 3 with GRPO

Learn to:
• Enable reasoning in Gemma 3 (1B)
• Prepare/understand reward functions
• Make GRPO work for tiny LLMs

Notebook: colab.research.google.com/github/unslo...
Details: huggingface.co/reasoning-co...

We made a Guide to teach you how to Fine-tune LLMs correctly!

Learn about:
• Choosing the right parameters & training method
• RL, GRPO, DPO, CPT
• Data prep, Overfitting, Evaluation
• Training with Unsloth & deploy on vLLM, Ollama, Open WebUI

🔗https://docs.unsloth.ai/get-started/fine-tuning-guide

Tutorial: Train your own Reasoning LLM for free!

Make Llama 3.1 (8B) have chain-of-thought with DeepSeek's GRPO algorithm. Unsloth enables 90% less VRAM use.

Learn about:
• Reward Functions + dataset prep
• Training on free Colab GPUs
• Running + Evaluating

Guide: docs.unsloth.ai/basics/reaso...

For our benchmarks, a standard GRPO QLoRA setup (TRL + FA2) for Llama 3.1 (8B) at 20K context required 510.8GB VRAM. Unsloth’s GRPO algorithms reduces this to just 54.3GB.

The 5GB VRAM requirement for Qwen2.5 (1.5B) is down from 7GB in our previous GRPO release two weeks ago!

Today, we’re launching new algorithms that enable 10x longer context lengths & 90% less VRAM for training Reasoning Models (GRPO).

Using Unsloth, you can now train your own reasoning model with just 5GB VRAM for Qwen2.5-1.5B with no accuracy loss.

Blog: unsloth.ai/blog/grpo

You can now reproduce DeepSeek-R1's reasoning on your own local device!

Experience the "Aha" moment with just 7GB VRAM.

Unsloth reduces GRPO training memory use by 80%.

15GB VRAM can transform Llama-3.1 (8B) & Phi-4 (14B) into reasoning models.

Blog: unsloth.ai/blog/r1-reas...

Finetune Phi-4 with Unsloth Fine-tune Microsoft's new Phi-4 model with Unsloth! We've also found & fixed 4 bugs in the model.

You can finetune Phi-4 for free on Colab now!

Unsloth finetunes LLMs 2x faster, with 70% less VRAM, 12x longer context - with no accuracy loss

Documentation: docs.unsloth.ai
We also fixed 4 bugs in Phi-4: unsloth.ai/blog/phi4

Phi-4 Colab: colab.research.google.com/github/unslo...

Introducing Unsloth Dynamic 4-bit Quantization!

Naive quantization often harms accuracy but we avoid quantizing certain parameters. This achieves higher accuracy using only <10% more VRAM than BnB 4bit

Read our Blog: unsloth.ai/blog/dynamic...
Quants on Hugging Face: huggingface.co/collections/...

Llama 3.2 Vision Fine-tuning with Unsloth Fine-tune Meta's Llama 3.2 Vision, Llava, Qwen 2.5 Vision models open-source 2x faster via Unsloth! Beginner friendly.

You can finetune Llama-3.2-Vision-11B for free on Colab now!

Unsloth finetunes VLMs 2x faster, with 50% less VRAM, 6x longer context - with no accuracy loss.

Other VLMs like Qwen2-VL, Llava & Pixtral are supported.

Blog: unsloth.ai/blog/vision

Fine-tuning Colabs: docs.unsloth.ai/get-started/...