How to fine-tune Qwen3.8 (LoRA, QLoRA)

This guide covers adapting models from the Qwen3.8 family. Memory and GPU counts appear below. See also LoRA, QLoRA, the training cost calculator and fine-tuning.

Licenses differ by model

Check the license before you build on a checkpoint. The model cards list:

• Qwen3.8-27B: Apache 2.0
• Qwen3.8-Flash-Next: qwen-community-1.0 (a custom community license)
• Qwen3.8-2.4T-A95B: qwen3.8-max (a custom license)

Review the full text of the two custom licenses for limits on derivatives and commercial use before training.

Which checkpoint to start from

The Qwen organization lists the three post-trained models and an FP8 variant of each. No Base checkpoints are listed. The FP8 build uses fine-grained FP8 quantization with block size 128, intended for serving. Train adapters on the unquantized repository (QLoRA quantizes on load).

Chat template and thinking

The models think by default. For Qwen3.8-27B and Flash-Next you can disable thinking with enable_thinking, and preserve_thinking keeps earlier reasoning in the history. Qwen3.8-2.4T-A95B always thinks, so training data for it should keep reasoning blocks. Use the shipped template in training and serving the same way, and decide whether your data carries reasoning traces.

Architecture notes

The 2.4T-A95B model is text-only and is a very large Mixture-of-Experts model, so full fine-tuning is rarely practical. The 27B model accepts images and video. The cards do not document LoRA target modules, so inspect module names in the checkpoint first.