MoonshotAI Ships Kimi K3, First Open 3T-Class MoE Model

MoonshotAI dropped weights and a technical report for Kimi K3, and the headline number is deliberately provocative. It is a 2.8 trillion parameter mixture of experts model, released with open weights, which Moonshot is calling the world's first open 3T-class system. On paper, that puts a Chinese lab's open release in the same weight class as the closed frontier coming out of the US labs.

The more interesting claim sits underneath the size. Rather than a straight scale-up of K2, the report describes two architectural moves called Kimi Delta Attention and Attention Residuals, paired with a Stable LatentMoE framework that activates only 16 of 896 experts per token. Together, Moonshot says these deliver roughly a 2.5x improvement in scaling efficiency over Kimi K2. That is the actual thesis: get more intelligence per unit of compute, not just more parameters.

On the scoreboard, VentureBeat reported K3 sitting third on GDPval-AA v2 behind Claude Fable 5 Max and GPT-5.6 Sol Max, and taking the top spot on a Frontend Code Arena with an Elo of 1,679. Moonshot's own launch material lists 93.5% on GPQA Diamond and 88.3% on Terminal-Bench 2.1. Take the specifics as reported, not settled, since several come from the vendor's own numbers and leaderboards move fast.

The honest caveat is that this is not a model most teams can casually run. Moonshot recommends a supernode with at least 64 accelerators for deployment, with MXFP4 weights and MXFP8 activations doing the memory heavy lifting. What the reporting does not give you is enough on the training data, the compute budget, or how the 2.5x efficiency claim was measured against K2 to fully audit it.

The forward-looking bit is the ecosystem release. Alongside the weights, Moonshot is opening high-performance attention kernels, an MoE communication library, and infrastructure for running agent environments at scale. If that stack is genuinely usable outside their own datacenters, it lowers the bar for anyone else trying to serve trillion-scale MoE models, and it is the piece I would watch more closely than the leaderboard row.