Moonshot AI has introduced Kimi K3, a new frontier-class open-weights model featuring 2.8 trillion total parameters and native multimodal input capabilities. Officially positioned as "Open Frontier Intelligence," the model supports a 1 million-token context window and utilizes novel architectural components like Kimi Delta Attention (KDA) and Attention Residuals to enhance efficiency.
- Kimi K3 employs LatentMoE with only 16 activated experts out of 896, achieving an activation ratio under 2%.
- The model claims up to 6.3x faster decoding in million-token contexts via KDA and ~25% higher training efficiency through Attention Residuals.
- Early evaluations place Kimi K3 at #1 in the Frontend Code Arena with a 76% pairwise win rate, surpassing Claude Fable 5 and GPT-5.6 Sol in that specific domain.
- Artificial Analysis ranks K3 comparable to Opus 4.8 and GPT-5.5 on their Intelligence Index, while noting its pricing is competitive with Sonnet 5.
- Open weights are promised for release by July 27, 2026, completing the transition from its current API availability.
The release marks a significant milestone in open-source AI by offering scale and performance previously reserved for closed models, though Moonshot acknowledges a remaining gap in user experience compared to top-tier proprietary systems.