Big model. 2.4T params, 95B active. 4.89TB. Surprised they didn't release a 4-bi

The 95-billion active-parameter figure describes only part of the model discussed in Alex Cheema's deployment post. Qwen's model card lists 2.4 trillion parameters in total and 95 billion activated, with a mixture-of-experts architecture that selects a subset of its experts for computation.

The linked release contains FP8 weights, which store numerical values at reduced precision. Cheema's proposed four-computer setup explicitly assumes a further reduction to four-bit weights. That makes it a proposal about a different representation of the model, rather than evidence that the linked release has been demonstrated on that setup.