BEIJING: The Kimi K3 launch introduced a 2.8-trillion-parameter, native multimodal model with a one-million-token context window, Chinese developer Moonshot AI said.
The company unveiled the model on July 14 and made it available through Kimi.com, Kimi Work, Kimi Code and its application programming interface. However, the full model weights are scheduled for release by July 27, according to Moonshot AI’s technical announcement.
Kimi K3 uses a mixture-of-experts architecture, activating 16 of its 896 experts for each token. Moonshot said the model also incorporates Kimi Delta Attention and Attention Residuals to improve efficiency and performance across long-context tasks.
The company has positioned Kimi K3 for software development, knowledge work, visual reasoning and complex, multi-step assignments. Its one-million-token context window allows the model to process considerably larger bodies of text and code within a single session.
At the time of publication, Arena’s Frontend Code leaderboard placed Kimi K3 first in that task-specific category. The result indicates strong frontend development performance, but by itself, does not establish overall superiority across artificial intelligence workloads.
Read: DeepSeek Valuation Hits $51.8bn in First Funding Round
Moonshot acknowledged that the model’s broader performance and user experience still trail leading proprietary systems, including Claude Fable 5 and GPT 5.6 Sol. It also identified excessive proactiveness and sensitivity to previous reasoning history among Kimi K3’s current limitations.
API access starts at $0.30 per million cached input tokens, rising to $3 for uncached input and $15 for output. Moonshot said additional technical documentation will accompany the full-weight release.
