NVIDIA is providing a quantization version of the autoregressive language model "Qwen3.8-27B" developed by Alibaba. This model has been quantization using NVIDIA's Model Optimizer with mixed precision combining NVFP4 and FP8.

As part of the quantization specifications, NVFP4 is applied to the MLP layers and the language model head (lm_head), while FP8 is applied to the self-attention and linear attention layers. This checkpoint can be deployed using vLLM. Additionally, NVIDIA Grace Blackwell GB300 is used as the test hardware.

The model supports text, image, and video inputs, with a maximum context length of 262,144 tokens. The output is in text format. Note that this model was not developed by NVIDIA, but was built based on third-party requirements.

Source: