Kimi K3: Open Frontier Intelligence
Kimi Team says it is releasing full weights for a 2.8T-parameter MoE model aimed at frontier-level open intelligence.
Kimi K3 activates 104 billion parameters per token, adds native vision, and supports a 1-million-token context window. The report attributes its scaling gains over Kimi K2 to Kimi Delta Attention, Attention Residuals, Stable LatentMoE, and revised training recipes. Its evaluations claim strong results in coding, agentic, knowledge, reasoning, long-context, and vision tasks, while still trailing the top proprietary models named in the abstract. Source: ArXiv · AI/CL/LG's note.
Kimi K3 activates 104 billion parameters per token, adds native vision, and supports a 1-million-token context window. The report attributes its scaling gains over Kimi K2 to Kimi Delta Attention, Attention Residuals, Stable LatentMoE, and revised training recipes. Its evaluations claim strong results in coding, agentic, knowledge, reasoning, long-context, and vision tasks, while still trailing the top proprietary models named in the abstract. Source: ArXiv · AI/CL/LG's note.
score 8