Moonshot releases Kimi K3.1 Pro with agentic gains and open weights slated for November 4

Moonshot launched Kimi K3.1 Pro as a multimodal mixture-of-experts model with 2.8 trillion total parameters, 104 billion active parameters, and a 1-million-token context window. The company describes it as a post-trained version of Kimi K3, with substantially larger post-training compute aimed at enabling large-scale joint reinforcement learning across hundreds of downstream domains.

Independent evaluations reported by Artificial Analysis placed K3.1 Pro fourth on its Intelligence Index with a score of 55, behind Opus 5.5 (58), Sonnet 5.5 (56), and Muse Wonder Full (56). Its strongest gains were in agentic tasks rather than uniformly distributed across knowledge and science tests. On Code Arena’s WebDev benchmark, K3.1 Pro ranked eighth with 1,692 points, up nine places and 37 points over Kimi K3, and Moonshot returned to the top four labs on Arena.

Moonshot also introduced a Highspeed API that preserves intelligence while generating output 2–3× faster, though it costs twice as much. The launch included Kimi Hosted Agents, a managed environment for long-running work and multi-agent collaboration with a 24/7 execution environment.

The open weights were not available at launch but are promised by November 4, 2026. The announcement did not specify the license, redistribution conditions, commercial-use restrictions, checkpoint precision, or complete inference requirements, so descriptions of the release as fully open source are premature.

Architecturally, the 104B active parameters do not equate to storing a 104B dense model; deployment must account for the much larger expert pool and runtime state. The posts do not disclose layer layout, expert routing details, or measured serving configurations.