Moonshot AI released its new-generation open-source model Kimi K3.1 Pro in the early hours of October 24. The model has a 3-trillion-parameter scale and supports a 1-million-token context window, remaining the world's largest open-source model by parameter count.
K3.1 Pro is a post-trained version of July's Kimi K3 (2.8 trillion parameters), with reinforcement-learning compute sharply increased so the model gets smarter across hundreds of domains. Gains are most notable in code, 3D creation, financial research and other long-horizon agent tasks: in multiple international evaluations its overall intelligence approaches leading closed models, taking global best scores on Office QA Pro v2 and SpreadsheetBench v2.
On Artificial Analysis's intelligence index, Kimi K3.1 Pro ranks fourth globally and first among open-source models - at a price just one-fifth of OpenAI's GPT-6 Astra. In one internal test, the model iterated on a simulated human-walking controller, fixing 36 shoulder-muscle moment-arm errors and extending continuous walking from a few stumbling steps to 11.2 seconds.
Moonshot said the complete model weights will be released before November 4. On safety, the team identified and removed cyber offense-defense data from mid-training and reinforcement learning, and pledged two open principles: not releasing weights that push the absolute frontier of cyber-attack capability while gradually increasing available defensive capabilities in the open weights.
Image: Efficiency of AI-related computer chips. Source: Wikimedia Commons (CC BY-SA 4.0).
