Moonshot AI, the Chinese AI startup, has unveiled Kimi K3, a 2.8-trillion-parameter model that the company describes as its most capable system to date. Released across the Kimi platform—including Kimi.com, Kimi Work, Kimi Code, and the Kimi API—the model represents a scaling achievement as the first open-weight system to approach the three-trillion-parameter threshold. Its full weights are scheduled for release by July 27, 2026.
Kimi K3 is built on two architectural innovations: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), which are designed to improve information flow across sequence length and model depth. The system employs a Mixture of Experts architecture with 896 experts, activating 16 per forward pass through the Stable LatentMoE framework. Moonshot AI reports that these structural changes yield 2.5 times greater scaling efficiency compared to its predecessor, Kimi K2.
The model features native multimodal capabilities, processing both text and images within a unified architecture, and supports a context window of one million tokens. In practical applications, Kimi K3 demonstrates strong performance in long-horizon software engineering, including autonomous GPU kernel optimization, compiler development, and end-to-end chip design using open-source EDA tools.
Beyond coding, the system handles knowledge work, scientific research synthesis, video editing, and interactive dashboard generation. The API is priced at $3.00 per million input tokens and $15.00 per million output tokens, with cached input discounted to $0.30.
Introducing Kimi K3: Open Frontier Intelligence
— Kimi.ai (@Kimi_Moonshot) July 16, 20262.8 Trillion Parameters, 1 Million Context, Native Multimodal
Independent evaluations indicate that Kimi K3 has quickly ascended the frontier model hierarchy. According to Arena.ai, the model claimed the top position in the Frontend Code Arena with 1,679 points, surpassing Anthropic’s Claude Fable 5 and marking a seventeen-place improvement over Kimi K2.6. It ranked first in six of seven frontend domains, including brand and marketing, data analytics, and content creation tools.
On the Artificial Analysis Intelligence Index, Kimi K3 scored 57, placing it in the same tier as Claude Opus 4.8 and GPT-5.5, though still trailing Claude Fable 5 and GPT-5.6 Sol. The model achieved an Elo rating of 1,668 on the GDPval v2 agentic task benchmark, outperforming GLM-5.2, GPT-5.5, and Claude Opus 4.8, while securing first place on AutomationBench-AA. In long-horizon knowledge work assessments, it reached 1,547 Elo on AA-Briefcase, a 732-point gain over K2.6 and second only to Fable 5.
Analysts note that Kimi K3 delivers this performance at a cost per task of $0.94, comparable to GPT-5.6 Sol and roughly half the cost of Opus 4.8. The model also improved token efficiency by 21% over its predecessor.
While some observers caution that early benchmarks may not fully reflect real-world reliability, the release has intensified discussion regarding the narrowing gap between Chinese and American AI capabilities.
The post Kimi K3 Ascends To AI Frontier: Chinese Open-Weight Model Challenges Top US Systems appeared first on Metaverse Post.
