Qwen3.8-27B Can Now Be Trained on a 24GB Graphics Card

Qwen is the model family Alibaba trains and releases directly. Qwen3.8-27B, whose weights were released under the Apache 2.0 license on August 14, shot to the top of Hugging Face's trending list right after launch, and in some user comparisons it reportedly outperformed Google's Gemini 3.7 Flash. That story was already covered in Qwen3.8-27B Runs on a Laptop, Beats Gemini 3.7 Flash. This latest announcement picks up where that left off — it's about the next stage, actually getting your hands on the released model and retraining it for your own use case.

What's New

Unsloth is a tool for fine-tuning already-trained models for specific purposes. According to the company's GitHub documentation, training Qwen3.8-27B with Unsloth runs about 1.5x faster than the standard FA2 (FlashAttention2) approach while using roughly half the VRAM. The company says this comes with no loss in accuracy, crediting the gains to a computation optimization technique called the Flash Linear Attention kernel.

Memory thresholds also vary by training method. QLoRA, which freezes most of the model's parameters and trains only a small portion at low cost, works with as little as 22GB. LoRA needs more than 36GB. Full fine-tuning (FFT), which retrains the entire model, requires roughly four times more VRAM than QLoRA, according to the training guide. Unsloth also supports reinforcement learning (RL)-based training, which it says can run for free or with 22GB of VRAM.

Getting Started

You don't need your own GPU to try this. Unsloth has opened an entry path through free notebooks on Kaggle, a Google-run service similar to Google Colab that offers free access to two Tesla T4 GPUs for 30 hours.

1. Open either Unsloth's Conversational notebook or its RL GRPO notebook for reinforcement learning.
2. In the Conversational notebook you can turn on vision support, while in the RL notebook you just need to set the model to Qwen3.8 and you're ready to train.
3. Load the 4-bit compressed Qwen3.8-27B-unsloth-bnb-4bit model and start training with QLoRA.
4. Once training is done, you can export the model in whatever format you need — NVFP4, FP8, or GGUF.

If you have your own GPU, you can download the Unsloth desktop app for macOS, Windows, or Linux and start training locally right away. For supervised fine-tuning (SFT) with text-only data, you'll need a dataset with a text column rendered using Qwen's chat template, as shown below.

python from unsloth import FastModel from datasets import load_dataset from trl import SFTTrainer, SFTConfig

max_seq_length = 2048 model, tokenizer = FastModel.from_pretrained( model_name = "unsloth/Qwen3.8-27B-unsloth-bnb-4bit", max_seq_length = max_seq_length, load_in_4bit = True, full_finetuning = False,

As an example use case, you could turn internal customer service logs into training data and run QLoRA in the Conversational notebook to build a chatbot tuned to a specific tone and body of business knowledge. If your data includes manuals with images, you could turn on the vision option to train a model specialized in document understanding.

Comparing What Each Method Requires

Training Method	VRAM Needed	Notes
QLoRA	22GB	Works even on a free Kaggle notebook
LoRA	More than 36GB	—
Full Fine-Tuning (FFT)	~4x QLoRA	For cases requiring high precision
RL (GRPO)	Free or 22GB	Reinforcement learning approach

Installation instructions and notebook links are available at the Unsloth GitHub repository.

Editor's Take

Even as open-weight models keep pouring out, only a small number of people have actually been able to get their hands on them and adapt them to their own data. The conventional wisdom was that properly fine-tuning a 27B-class model required data center-grade GPUs like the A100 or H100. This announcement matters because it pushes that baseline all the way down to a single gaming card — or even a free cloud GPU. Behind Qwen3.8-27B's rise to the top of Hugging Face's trending list and its reported edge over Gemini 3.7 Flash lies the fact that training tools individual developers can pick up and use right away arrived alongside it.

Working with fine-tuning tools for models this size tends to lead to the same conclusion every time: start with a lightweight method like QLoRA to check whether you're heading in the right direction, then move to full fine-tuning if needed. That sequence saves the most time and GPU cost. For a Korean startup or an in-house AI team, it makes sense to start by running QLoRA experiments on internal data. Thirty free hours on a Kaggle notebook is plenty of time to complete one small-scale dataset experiment, and if the results look promising, there's no rush — you can consider investing in your own GPU or renting cloud compute afterward.

That said, it's still too early to jump straight into retraining the entire model, especially one that supports a 262K context window along with vision and video input. FFT uses four times the VRAM, so it makes sense to first validate at the QLoRA stage whether that level of investment is even worth it. In the coming weeks, expect to see a number of derivative versions of this model, retrained for specific tasks, showing up on Hugging Face. As open-weight models and lightweight training tools continue to reinforce each other, the next round of competition will likely be among tool makers racing to make this combination the standard development environment.