Moonshot AI releases open-source model Kimi K3.1 Pro, 3 trillion parameters, 1 million token context

Beijing News Shell Finance Report (Reporter Zhang Xiaohui) In the early hours of October 24, Moonshot AI released a new-generation open-source model, Kimi K3.1 Pro.

It is understood that the model has a parameter scale of 3 trillion and supports a 1 million token context window, making it the open-source model with the largest parameter scale in the world at present. Moonshot AI stated that the complete model weights will be released before November 4.

Moonshot AI introduced that Kimi K3.1 Pro has increased compute investment in the reinforcement learning phase, allowing the model to become smarter through repeated training across hundreds of domains; on the basis of possessing an intelligent "brain," it has grown more "hands," with stronger professional tool-calling capabilities, and has improved in code writing, 3D creation, financial research and other long-horizon Agent capabilities, and can complete longer multi-step complex tasks.

In terms of security, Moonshot AI introduced that the R&D team conducted a special assessment of Kimi K3.1 Pro's vulnerability research capabilities and implemented targeted controls during training. In addition, the team identified and removed data related to cyber offense and defense from the mid-training and reinforcement learning stages, and will continue to adopt such controls in each subsequent training stage. Assessments show that although Kimi K3.1 Pro's cyber offense and defense capabilities have improved compared with the previous generation, they still lag behind frontier closed-source models that have been publicly released for at least three months, do not exceed the scope of attackers' existing capabilities, and the risk is within controllable range.

At the same time, the Moonshot AI research team clarified two openness principles: not releasing open weights that advance the absolute frontier of cyber offense and defense capabilities, and gradually improving the cyber offense and defense capabilities available in open weights. The team believes that opening the K3.1 Pro model weights will be more beneficial to groups that need offense and defense capabilities for defensive purposes. When capable models are provided in open form and can run on user-controlled infrastructure, more teams will be able to review the code they maintain. Letting this capability gradually diffuse, rather than concentrating in a few closed-source systems, will ultimately make all software more secure.

1 Livestream | A magnitude 7.0 earthquake occurred in Baqing, Tibet; live coverage of the latest rescue progress in the quake zone 128,000

2 Wang Zhijun appointed Secretary of the Hunan Provincial Party Committee 103,000

3 Epicenter 10 kilometers deep, visual investigation of rescue in the magnitude 7.0 earthquake in Baqing, Tibet | Rebuilding the scene 101,000

4 The new Henan Provincial Party Committee leadership team appears 83,000

5 Low-price bubble fades, how can merchants achieve growth with more substance? Meituan Waimai conference solves the problem 53,000

Moonshot AI releases new-generation large model Kimi K3.1 Pro 35 minutes ago

Zhuge Yujie appointed Party Secretary of the Qinghai Provincial Government 41 minutes ago

Hot discussion

"One month after the start of high school, every class I teach has students taking leave due to mental distress" | Quick comment

Life rescue on the northern Tibetan plateau—24-hour live coverage of rescue in the Baqing, Tibet earthquake

Health departments carry out medical emergency work for the earthquake in Baqing County, Tibet

Red leaves in the Wucai Qianshan area of Shunyi enter the best viewing period, lasting until mid-November

Beijing releases 4 Long March-themed teaching resources, 50,000 students will visit the Military Museum for exhibitions