You can now feed an entire book-length context into a single frontier model and tell it to reason more deeply or more quickly.
Moonshot AI put that capability into a new family called Kimi K3.1 Pro, with a faster Highspeed version. They are available across Kimi Code, Kimi Work, Moonshot's web tools and the API. The company said it will publish the complete model weights in early November. Reuters independently corroborated the launch and the planned weight release.
That matters because long-memory plus controllable reasoning changes how teams handle large, complex documents. Tasks that used to require chopping a file into many short queries can now be done in a single pass: contract review, long code reasoning, or multimodal reports become simpler to build and maintain.
The model can take in an extremely long input at once and offers "thinking" modes labelled low, high and max. Think of it like reading the whole book and choosing whether to skim, study, or interrogate it carefully. There is also a Highspeed sibling meant to produce answers noticeably faster at similar quality. Moonshot says it removed cyber-related data from training and reinforcement learning to lower offensive risks, and Reuters reports their internal tests show a measurable drop on that cyber benchmark after the change.
In practice, companies and researchers can start building workflows that span huge documents and multi-step reasoning without stitching together many short-context calls. Publishing weights does not mean everyone can run the model on a laptop: expect enterprise-scale GPU infrastructure or cloud instances to be required to use the raw weights.
The real test is whether independent teams can turn these published blueprints into products that compete with closed, hosted systems and whether regulators will respond to the vendor's safety trade-offs.
