On July 16, 2026, China’s Moonshot AI officially released its flagship large model KIMI K3, quickly becoming a focal point in China’s domestic AI field. The model, with approximately 2.8 trillion parameters, a 1,048,576-token (1M) context window and native multimodal capabilities, set multiple records for China-made large models. In independent benchmarks such as Artificial Analysis, the KIMI K3 Intelligence Index reached 57, ranking among the leaders and demonstrating competitiveness with international state-of-the-art models. However, despite eye-catching benchmark results, real-world performance still requires further market validation.

KIMI K3 core specifications and benchmark results

KIMI K3 uses a mixture-of-experts (MoE) architecture with around 2.8 trillion parameters, a context window of up to 1,048,576 tokens (1M), and supports text and image inputs with text output. The architecture introduces innovations such as Kimi Delta Attention and Attention Residuals to improve long-sequence processing efficiency and information transmission.

Key benchmark results (based on independent benchmarks including Artificial Analysis):

Key benchmark results (based on the latest independent benchmarks, maxed out thinking effort):

On Terminal Bench 2.1, KIMI K3 scored 88.3, just behind GPT-5.6 Sol (88.8), and substantially ahead of Opus-4.8 and Fable 5 (both 84.6), showing strong interactive and execution capabilities.
On Program Bench, KIMI K3 ranked first with 77.8, slightly above GPT-5.6 Sol (77.6), demonstrating strong program generation and logical reasoning performance.
On the FrontierSWE benchmark, KIMI K3 scored 81.2, behind Fable 5 (86.6) but well ahead of GPT-5.6 Sol (71.3), highlighting competitiveness on advanced software engineering tasks.
On SWE Marathon, KIMI K3 led with 42.0, showing advantages in long-sequence encoding and handling complex projects.
On DeepSWE and Kimi Code Bench 2.0, KIMI K3 also held leading positions; overall coding ability approaches or surpasses some international flagship models.

Moonshot’s official statement notes that K3 shows clear advantages in long-horizon agent and complex workflow scenarios, with pricing at input $3 per million tokens and output $15 per million tokens (cache hits $0.30), offering notable cost-performance. Model weights are planned to be open-sourced on July 27, continuing the KIMI series’ open strategy.

Technical innovations behind the performance highlights

The strong benchmark performance of KIMI K3 reflects several technical breakthroughs:

Massive MoE architecture: 2.8 trillion parameters use expert routing for efficient activation, balancing performance and inference cost.
Long-context optimization: The 1M window combined with new attention mechanisms addresses efficiency bottlenecks of traditional Transformers on long sequences.
Multimodal fusion: Native visual capabilities give the model advantages on mixed image-and-text tasks.
Agentic capabilities: Supports complex workflow planning, suitable for enterprise-grade applications.

Compared with Chinese domestic competitors such as DeepSeek V4 Pro, Wenxin 5.0 and Tongyi Qianwen 3.7 Max, KIMI K3 leads in parameter scale and context length, but its real-world deployment effects still require time to verify.

Potential impacts: domestic AI ecosystem and global competition

1. Impact on Moonshot

Brand and valuation boost