Kimi-K3 Releases on HuggingFace 7/27

What commenters are saying

Commenters focused on the economics of hosting a 3T-parameter mxfp4-native model requiring ~1.5TB VRAM, likely needing 16x B200s for practical use. They debated whether third-party inference pricing will reveal if top labs subsidize API costs, with some pointing to estimates that Anthropic's blended gross margins are in the mid-60% range. Others noted that inference efficiency gains from RL training complicate cost analysis. Some discussed running the model on CPU servers with 3TB RAM for slow but cheaper inference, while others argued electricity costs would outweigh API savings.