Moonshot AI Releases Kimi K3.1 Pro: A 2.8T Total/104B Active Params (MoE) Model with 1M Context

Why it matters

Moonshot AI has launched Kimi K3.1 Pro, the first model in its K3.1 series and a post-trained version of Kimi K3 built with substantially more post-training compute. The company ran joint reinforcement learning across a wide range of downstream domains, supported by work on its RL framework, long-context inference, sandbox infrastructure, and rollout scheduling. The model offers a very long context window, accepts text, images, and video, and is available now through the Kimi API and related products, with open weights expected later. Self-reported benchmarks show it leading on a software engineering marathon test while trailing GPT-6 Astra and other proprietary rivals on science-heavy terminal tasks. Moonshot also removed cyber-related data from mid-training and reinforcement learning ahead of the open-weight release, and says the shipped model stays behind frontier proprietary systems on offensive cyber tasks. Notably, several claimed wins come from internal benchmarks, and the model's size means it requires supernode-class serving rather than single-GPU deployment.