Kimi K3's 2.8 Trillion Open Weights Just Landed — Only 1.8% of Them Ever Fire

Moonshot AI released the Kimi K3 open-weight model with 2.8 trillion parameters on Hugging Face on July 27, but only 1.8% of parameters (about 50 billion) are active per forward pass due to Mixture-of-Experts routing. The model requires 1,488 GB of VRAM, exceeding the capacity of an 8× H200 node (360 GB short) and an 8× B200 node (48 GB short), contradicting self-hosting guides.

Member-only story Kimi K3's 2.8 Trillion Open Weights Just Landed — Only 1.8% of Them Ever Fire Moonshot AI pushed the Kimi K3 weights to Hugging Face at 00:00 UTC on July 27, and the number everyone quoted was 2.8 trillion. It is the largest open-weight model ever released, by a wide margin. The number nobody put in the headline is 1.8%. Kimi K3 routes each token to 16 of its 896 experts. That is 1.786% of the expert pool, and roughly 50 billion of the 2.8 trillion parameters doing any actual work on a given forward pass. The other 98.2% sit in GPU memory, paid for, idle, waiting for a token whose router score picks them. I spent the morning after the drop doing the arithmetic that the launch posts skipped: what does 2.8 trillion parameters at 1.8% sparsity actually cost you in VRAM, and which box can hold it? The answer is 1,488 GB — and it produced a result I did not expect. Half the “how to self-host Kimi K3” guides published in the last 48 hours tell you an 8× H200 node will do it. That node is 360 GB short. An 8× B200 node, the one that looks like it fits, is short by 48 GB. Here is the math, the script that reproduces it, and the license clause that makes “open weights” and “open source” two different things.