- Moonshot AI's Kimi K3.1 model identifier was found callable in the company's API backend by multiple developers on September 28, before any official announcement.
- The leaked configuration shows three adjustable reasoning intensity levels — Low, High, and Max — suggesting tiered compute allocation and pricing.
- K3.1 may natively support Agent mode and Swarm multi-agent collaboration alongside search and batch processing task modes.
- API pricing currently mirrors the existing K3 model but is likely a placeholder ahead of an official launch.
Backend Discovery Precedes Official Preview
On September 28, multiple developers independently discovered a 'kimi-k3-1' identifier in Moonshot AI's API platform backend model registry. The identifier was not merely listed but could be successfully invoked through API interface probing, confirming the model was deployed and responsive.
Later that same day, Moonshot's official Kimi open platform published a preview listing for K3.1, identified internally by the codename 'K311111,' specifying support for a 1-million-token context window. The sequence — backend discovery followed by an official platform teaser — suggests the model is near public release.
Tiered Reasoning and Potential Agent Capabilities
The leaked configuration reveals three adjustable reasoning intensity levels: Low, High, and Max. Developers analyzing the setup interpret this as a mechanism for matching compute resources and billing to task complexity, allowing users to select lower-cost inference for routine queries or maximum reasoning depth for demanding workloads.
Beyond tiered reasoning, the configuration hints that the model may natively support Agent mode for autonomous task execution and Swarm multi-agent collaboration, where multiple model instances coordinate on complex problems. The same leaked configuration also suggests the model may support additional task modes such as web search and batch processing, though these remain unconfirmed ahead of an official announcement.
Pricing Placeholder and Competitive Implications
The Kimi open platform currently displays K3.1 API pricing identical to the existing K3 model. Given that K3.1 has not been officially announced, this pricing is widely expected to be a placeholder subject to change at launch.
K3.1's differentiation appears to lie not in context length — which matches the 1-million-token window already available on K3 — but in its configurable reasoning architecture and potential agentic capabilities. This reflects a broader shift in China's large model competition from conversational benchmarks toward autonomous task completion, as the industry moves from asking 'who chats better' to 'who can get work done independently.'
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01
