Moonshot AI, the Beijing-based artificial intelligence company, has officially launched Kimi K3, its most expansive language model to date. The model contains 2.8 trillion parameters, and the company describes it as the largest AI system with weights that are scheduled to be publicly released. Access is immediate and straightforward: K3 is available now through the Kimi app for iOS and Android, the kimi.com web interface, and the Kimi Work desktop client. A free account is sufficient to begin interacting with the model, though Moonshot notes that precise usage limits are adjusted dynamically within the application.
Self-Assessed Benchmarks and Early Independent Results
In an unusually candid move, the company has published its own benchmark comparisons that show K3 lagging behind Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol, while outperforming Claude Opus 4.8 and GPT-5.5 among other competitors. Early third-party data aligns with this picture. The evaluation platform Artificial Analysis currently places K3 directly behind Fable 5 in its private assessments. Additionally, K3 has secured the top position on Arena.ai’s frontend coding leaderboard. Most other available metrics still originate from Moonshot’s internal testing.
Architecture, Context Window, and Availability Timeline
On a technical level, Kimi K3 employs a mixture-of-experts architecture. For each individual request, the model activates 16 out of a total of 896 specialized expert sub-networks. It supports a context window of 1 million tokens and can process images natively. The predecessor model, Kimi K2, had already attracted attention for delivering strong benchmark performance. For users who wish to deploy K3 on their own infrastructure, the open weights are not yet published. Moonshot has committed to releasing them by July 27. Until that date, the model is accessible exclusively through the company’s own applications, website, and application programming interface.
Pricing Structure and Operational Behavior
API access introduces a new pricing tier for Chinese-language models. Moonshot charges $3 per million input tokens and $15 per million output tokens, positioning K3 above any previous domestic offering in cost. The model is currently configured to operate at maximum computational effort on every query. This design choice produces unusually thorough answers but results in noticeably slower response times. The company has indicated that lighter, faster operating modes will be introduced in a later update. As with any cloud-based service, users testing K3 are sending their prompts to servers operated by a Chinese provider, and standard caution applies to sensitive content. Moonshot also acknowledges that in overall user experience, K3 still trails GPT-5.6 Sol and Claude Fable 5. Even so, for anyone seeking an accessible gauge of China’s accelerating AI development, it represents the most compelling option available today.
Sources: www.kimi.com, simonwillison.net