Transformers 5.16 lands NVFP4 4-bit: run big models in half the VRAM
Day-one support lands too for Qwen4-Exp, Step-3.7-Flash and GLM-5.3-Flash, so the newest open models load straight from the standard library.
Copy markdown
NVFP4 halves your weight memory
transformers 5.16 wires NVFP4 4-bit quantization through Hugging Face kernels — on-the-fly BF16 weight quantization for ~50% less weight memory, often the line between a checkpoint fitting your GPU or not. Prequantized NVFP4 weights are already up, including nvidia's 397B Qwen.
Four new architectures, day one
v5.16.0 adds Qwen4-Exp (hybrid sparse + linear attention), Step-3.7-Flash (StepFun's 198B sparse-MoE vision-language model, ~11B active per token, Apache 2.0), Granite Speech 5.0 Turbo CTC (470M ASR), and CohereCompass — each loadable with the usual from_pretrained.
GLM-5.3-Flash multimodal in 5.16.1
The 5.16.1 patch adds GLM-5.3-Flash, the first natively multimodal model in the GLM-5 line (320B total / 18B active), so last week's MIT-licensed weights now run from mainline transformers instead of a fork.