Moonshot releases Kimi K3, a 2.8T-parameter open MoE model

A Chinese lab just dropped what is reportedly the largest open-weight model released to date, and the more interesting part is what Moonshot claims about efficiency rather than size. The Kimi K3 technical report puts the model at 2.8 trillion parameters in a mixture-of-experts configuration that activates 16 of 896 experts per token, with native vision and a roughly one-million-token context window.

The pitch on X leans on architecture rather than raw scale. Moonshot describes the design as delivering '2.5x the intelligence per unit of compute, not just more params.' The named building blocks are Kimi Delta Attention, a hybrid linear attention mechanism the company reports delivers up to 6.3x faster decoding in million-token contexts, and a Stable LatentMoE framework that pushes sparsity further than the previous Kimi K2.

The benchmark picture is genuinely mixed and worth taking as reported, not settled. According to Tom's Hardware, K3 landed first on Arena's Frontend Code evaluation at 1,679 points, ahead of Claude Fable 5, while sitting fourth on Artificial Analysis's Intelligence Index with a score of 57, behind Fable 5 and two GPT-5.6 Sol reasoning settings. Priced at $3 per million input tokens and $15 per million output tokens on the hosted API, it is positioned as a serious frontier alternative rather than a cheap-and-cheerful open release.

Moonshot is also opening up parts of the stack around the model: the attention kernels, an MoE communication library, and infrastructure for running agent environments at scale. That is arguably more consequential than any single benchmark line if you are trying to build on top of the weights. The honest caveat is that a 2.8T-parameter model is not something most teams will self-host, so the near-term practical win is downstream distillation, fine-tuning, and research access to the recipe. What the reporting does not yet answer is how the training compute was assembled under current export restrictions, or how K3 performs on multimodal evals outside the headline scores. For anyone running a make-versus-buy conversation on frontier capability, the ceiling for open weights just moved.