Kimi K3 Drops Today: The World's Largest Open-Weights Model at 2.8 Trillion Parameters

Kimi K3 open weights land today: 2.8T-parameter sparse MoE at 1.4TB quantized, needs 8x H100 minimum -- impractical to self-host for most, but it sets a new open-weight frontier that gets distilled fast.

Moonshot AI drops the full Kimi K3 weights on July 27. With 2.8 trillion total parameters -- 896 experts trained, 16 active per token -- it's the world's first open-weight 3T-class model. Architectural highlights: Kimi Delta Attention (KDA) for efficient long-sequence scaling, and Stable LatentMoE yielding 2.5x better scaling efficiency than Kimi K2.

The catch is hardware: MXFP4 weights still require ~1.4TB of GPU memory. Minimum viable setup is 8x H100 80GB; a well-equipped local workstation can't load it. The practical path for most engineers is the Kimi API or waiting for community-derived distillations -- models at this scale typically spawn runnable 7B-70B derivatives within weeks.

Why it matters: Moonshot has set the open-weight size frontier for nine of the past twelve months. Kimi K3 will drive the next wave of open-source fine-tunes and distillations. If you're planning a fine-tune on the strongest possible open base, or watching the OSS frontier for your production stack, this is the model to track -- even if you can't run the full weights yourself.

Sources: Kimi K3 Tech Blog · Kimi K3 model overview on HuggingFace · The open-weights escalation -- Interconnects.ai