Moonshot AI's next flagship model surfaced the way many things do now: developers found it first. On September 28, several developers noticed a "kimi-k3-1" identifier in the backend model registry of Moonshot's API platform — not merely listed, but passing interface probes and actually callable. Later that evening, the official Kimi Open Platform posted a preview page for K3.1 under the codename "K311111," stating support for a context window of up to 1 million tokens.
Based on the leaked configuration, K3.1 offers three adjustable reasoning-effort levels — Low, High and Max — and may natively introduce an Agent mode with Swarm multi-agent collaboration, alongside search and batch-processing task modes. The official platform currently lists K3.1's API pricing as identical to the existing K3, though since the model has not been formally announced, that figure is likely a placeholder rather than a final number.
Developers reading the configuration have converged on the same interpretation: splitting reasoning intensity into multiple tiers almost certainly maps to different compute allocations and billing models. A user can pick Low for cheap, high-volume work and Max for quality-critical tasks — an upgrade, as one analysis put it, from a single model working alone to a model commanding a group of models.
The predecessor context matters. Kimi K3 shipped with 2.8 trillion parameters and a 1 million-token context window, supporting visual understanding and long-horizon coding with configurable reasoning effort. K3.1 appearing on top of that base — pitched around agent-native modes rather than a raw capability bump — signals where Moonshot thinks the next round of competition will be won.
It also fits a broader pattern in China's model race. Xiaomi's MiMo-V2.6-Pro currently tops the open-weights rankings on the Artificial Analysis index, Alibaba is pushing agent infrastructure through its Qwen stack, and DeepSeek keeps competing on inference price. Across the board, the target has shifted from who chats better to who can complete tasks on their own — and Moonshot, which has been reported to be preparing a Hong Kong IPO, needs an agentic flagship to stay in that conversation.
Nothing is officially confirmed beyond the preview page: the leaked specs could change before launch, and Moonshot has not announced a release date. But the sequence — backend identifier spotted, interface probe succeeds, official teaser goes up the same night — is the customary on-ramp to a release. When K3.1 lands, the pricing page will be the number to watch: whether Max-tier reasoning is priced like a premium product or like a commodity will say a lot about where China's model price war goes next.
Comments (0)
Log in to join the discussion
Log InNo comments yet