Qwen 2.5 72B

Qwen 2.5 72B is the large member of Alibaba Cloud’s Qwen 2.5 autoregressive language-model family. The cited checkpoint is a pretrained base model for causal text generation rather than an instruction-tuned chat model. The memorable contrast is with Qwen 2.5 7B: both share a generation name, but 72B is roughly ten times larger by parameter count and therefore needs a larger deployment. It is still a dense language model, not a Mixtral-style sparse mixture of experts. TT-Metal lists Qwen 2.5 72B with TP=8 on a Wormhole QuietBox at batch 32. Tensor parallelism splits its tensors across eight chips, while the smaller 7B entry uses TP=2 on n300.