China’s Kimi K3 rivals the world’s biggest AI models and it’s open source

Moonshot AI, the Beijing based startup backed by Alibaba, has released Kimi K3, a 2.8 trillion parameter model the company calls in an official blog post the world’s first open 3T-class model. The system combines two new architectural components, Kimi Delta Attention and Attention Residuals, with a 1 million token context window and native vision capabilities.

According to Moonshot’s own benchmark data, Kimi K3 still trails the strongest proprietary systems, Claude Fable 5 and GPT 5.6 Sol, but performs at a frontier level across coding, reasoning and agentic tasks. Michael Nuñez reports for VentureBeat that independent evaluations from analytics firm Artificial Analysis largely confirm this picture: K3 placed third on the GDPval-AA v2 benchmark for real world professional tasks and second on the AA-Briefcase test for long horizon knowledge work, trailing only Fable 5 in both cases.

On BrowseComp, a benchmark for difficult, long horizon web research, K3 reached a score of 91.2, the highest among all tested models, achieved with a single agent and no context compression despite the model’s massive one million token window.

Chip design and scientific research as showcases

Moonshot AI used several case studies to demonstrate long horizon agentic ability. In one, K3 spent 48 hours autonomously designing, optimizing and verifying a small computer chip meant to run a compact version of itself, using open source design tools. The resulting chip covers just 4 square millimeters and reaches over 8,700 tokens per second in simulation.

In another example, the model reproduced a set of complex astrophysics equations, known as the I-Love-Q relations, cross checking more than 20 research papers and writing over 3,000 lines of code in about two hours, a task the company says normally takes a researcher one to two weeks.

A comeback after losing ground to DeepSeek

The release also marks a turnaround for Moonshot AI. Nuñez notes that the company lost significant market share after the launch of DeepSeek’s R1 model in early 2025, slipping from third to seventh place among Chinese AI chatbots by user numbers. Its pivot toward open source models, starting with Kimi K2 last July, appears to have paid off: Moonshot’s own comparison chart puts K3 far ahead of other open models by parameter count, including DeepSeek’s V4 Pro at roughly 1.6 trillion parameters.

Releasing such a large model as open source also carries strategic weight. Nuñez points out that Chinese companies increasingly use open source releases to expand global developer influence and to counter Western restrictions on advanced chips.

Availability and pricing

Kimi K3 is already accessible through the Kimi app, the Kimi Work desktop tool, the Kimi Code terminal agent and the Kimi API, where it costs 3 dollars per million input tokens and 15 dollars per million output tokens, with cached inputs priced at just 0.30 dollars. Full model weights will follow on July 27, according to Moonshot’s own announcement.

Sources

• Kimi K3 Tech Blog: Open Frontier Intelligence – Kimi Blog
• China’s Moonshot AI releases Kimi K3, the largest open-source model ever, rivaling top U.S. systems – VentureBeat