Qwen3.8

Qwen3.8-27B is Qwen’s dense 27B native vision-language model for text, image, and video inputs. It retains the Qwen3.5 hybrid Gated DeltaNet and gated-attention architecture, so NeMo AutoModel uses its existing native Qwen3_5ForConditionalGeneration implementation for the checkpoint.

Example Models and Recipes

Run the Recipes

$
uv run automodel --nproc-per-node=8 examples/vlm_finetune/qwen3_8/qwen3_8_27b.yaml
$
uv run automodel --nproc-per-node=8 examples/vlm_finetune/qwen3_8/qwen3_8_27b_lora.yaml

Install the vlm-media extra before running either recipe. See the Installation Guide and the Qwen3.8 fine-tuning guide for setup, validation details, and measured training-loss curves.

The checked-in recipes use FSDP2 data parallelism on eight GPUs. The shared Qwen3.5 dense model implementation also supports TP, CP, and PP, but the Qwen3.8 runs documented here validate the DP8 recipes specifically.