Kimi K3: Open Frontier Intelligence
Kimi’s team says K3 is a 2.8T-parameter open MoE model with released weights and frontier-level benchmark results.
The model activates 104B parameters per token, supports native vision, and carries a 1-million-token context window. Its report credits Kimi Delta Attention, Attention Residuals, Stable LatentMoE, and new training infrastructure for about a 2.5x scaling-efficiency gain over Kimi K2. The authors say K3 still trails Claude Fable 5 and GPT-5.6 Sol overall, but beats the other open and proprietary models in their evaluation suite. HF Daily Papers' note
The model activates 104B parameters per token, supports native vision, and carries a 1-million-token context window. Its report credits Kimi Delta Attention, Attention Residuals, Stable LatentMoE, and new training infrastructure for about a 2.5x scaling-efficiency gain over Kimi K2. The authors say K3 still trails Claude Fable 5 and GPT-5.6 Sol overall, but beats the other open and proprietary models in their evaluation suite. HF Daily Papers' note
score 8