Moonshot AI plans to release Kimi K3's full model weights on July 27. The model is already available through the company's Kimi platform, while API access is priced at $3 per million input tokens and $15 per million output tokens. Cached input tokens are priced at $0.30 per million, and the company is offering a temporary rebate program for qualifying API credit purchases through August 12.
A Moonshot AI executive, quoted by Xinhua, described the model's scale by comparing parameters to neural connections in the human brain, saying that nearly three trillion parameters allow the model to "store more knowledge and patterns in its brain, understand more, think deeper, and answer more accurately."
According to benchmark results shared by Moonshot AI and analytics firm Artificial Analysis, Kimi K3 ranks among the strongest AI models currently available. On the GDPval-AA v2 benchmark, which evaluates real-world tasks across multiple industries and occupations, the model placed behind only Claude Fable 5 Max and GPT-5.6 Sol Max. On Artificial Analysis' AA-Briefcase benchmark for long-horizon knowledge work, Kimi K3 ranked second overall, outperforming GPT-5.6 Sol Max. The company also reported a score of 91.2 on BrowseComp, a benchmark focused on complex information-seeking tasks.
Moonshot AI said Kimi K3 achieved these results using a single-agent architecture powered by its full 1 million-token context window, without relying on context compression or additional context management techniques.
In coding and automation benchmarks, the company said Kimi K3 ranked among the top-performing models across several evaluations, including first-place finishes on Automation Bench, SpreadsheetBench 2, BrowseComp, SWE Marathon, and Program Bench. The model also claimed the top position on Arena.AI's Frontend Code Arena leaderboard, which measures human preference in head-to-head frontend coding comparisons.
Beyond benchmark performance, Moonshot AI highlighted several demonstrations intended to showcase Kimi K3's autonomous agent capabilities. In one experiment, the model spent 48 hours independently designing a physical chip capable of running a smaller version of itself using open-source electronic design automation software. According to the company, the process covered architectural design, optimization, verification, and simulation, resulting in a functional 4-square-millimeter chip design that achieved timing convergence at 100 MHz and simulated decoding speeds exceeding 8,700 tokens per second.
The company also reported that Kimi K3 reproduced the universal I-Love-Q relation in computational astrophysics by reading more than 20 research papers, validating results, and implementing a complete numerical pipeline in approximately two hours. Moonshot AI said the same work would typically require one to two weeks for an experienced researcher.
The release represents a significant milestone for Moonshot AI, which rose to prominence after launching its Kimi chatbot before losing market share following DeepSeek's emergence. The company later shifted its strategy toward open-source AI models with the releases of Kimi K2 and K2.5, leading to Kimi K3 as its latest flagship system.
Ahead of the launch, the Financial Times reported that Kimi K3 was expected to perform at or above Anthropic's Opus 4.8 while becoming China's largest open-weight AI model, with a parameter count between 2 trillion and 3 trillion. The publication also reported that Moonshot AI was seeking additional funding at a higher valuation following previous fundraising rounds.
Alongside Kimi K3, Moonshot AI released updates to its open-source Kimi Code coding assistant, adding expanded subagent tools, background task management, security improvements, nested agents, and planning features. The company said the tool integrates with development environments including VS Code, Cursor, and Zed.
Moonshot AI's current model lineup includes Kimi K3 as its flagship model, K2.7 Code for software development tasks, and K2.6 as a general-purpose model. All three support context windows of at least 256,000 tokens, while K3 extends that limit to 1 million tokens. The company also said context caching operates automatically without requiring developers to configure additional parameters.
The release comes as Chinese AI companies continue expanding their open-source offerings alongside rivals including DeepSeek, Alibaba, Tencent, and Baidu. By open-sourcing what it says is the world's largest AI model, Moonshot AI is positioning Kimi K3 as a platform for developers and enterprises seeking alternatives to proprietary AI systems while competing directly with frontier models from OpenAI and Anthropic.
This analysis is based on reporting from Venture Beat.
Image courtesy of Kimi AI.
This article was generated with AI assistance and reviewed for accuracy and quality.
