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

Moonshot AI just dropped Kimi K3 on Hugging Face, a 2.8-trillion-parameter mixture-of-experts model with a million-token context window and native multimodal input. The interesting part isn't the raw parameter count, it's that the weights are actually out in the open.

The architecture is the shape worth looking at. Only 104 billion of those 2.8 trillion parameters activate per token, drawn from a pool of 896 experts with 16 selected at a time. Moonshot pairs that sparsity with a new attention mechanism they call Kimi Delta Attention and a framework, Stable LatentMoE, that they claim delivers roughly 2.5x scaling efficiency over Kimi K2. Vision is handled by a 401M-parameter MoonViT-V2 encoder inside the same stack, and the model ships quantized to MXFP4 weights and MXFP8 activations.

On benchmarks, Moonshot positions K3 against Claude Opus 4.8, Claude Fable 5, GPT-5.5, GPT-5.6 Sol and GLM-5.2. The self-reported numbers include 93.5 on GPQA Diamond, 88.3 on Terminal-Bench 2.1 and 91.2 on BrowseComp. According to reporting on the launch cadence, K3 first shipped through Kimi.com and the Kimi API on July 17, with the open weights following on July 26, a day ahead of the July 27 target.

Take the specifics as reported, not settled. These are vendor-published charts, and even Moonshot's own slides have K3 trailing Fable 5 on FrontierSWE and GDPval-AA Elo and GPT-5.6 Sol on DeepSWE. What the reporting doesn't give you is training-data provenance, real serving cost per token at MXFP4, or how Kimi Delta Attention actually behaves at the far end of that million-token window in production workloads.

The downstream question isn't whether Kimi tops any single leaderboard. It's what happens when a frontier-tier open weight model with a real coding and agent story lands in the hands of teams that were stuck choosing between paid API bills and second-tier open models. The pricing pressure on hosted frontier vendors gets one notch harder to ignore, and inference platforms suddenly have a new headline model to package.