❯ Moonshot releases Kimi K3.1 Pro weights as Moore Threads adds GPU support

Weights released for a three-trillion-class modelMoonshot released the full weights of Kimi K3.1 Pro on November 4. The first model in the K3.1 family has roughly three trillion parameters and a one-million-token context window. It supports coding, 3D creation, financial research and multi-step agent tasks. Developers can use Kimi products and APIs or download the weights for their own deployments.

Moore Threads runs basic inferenceMoore Threads announced its Day-0 adaptation on November 6. MTT S5000 cards and the MUSA software stack support model loading, service startup and basic text and image inference, adding a Chinese GPU option for deployment.

Software covers the key computationsThe adaptation uses SGLang-MUSA, operator libraries and compilers to support sparse attention, cache management and vision computation. Released weights still require compatible software to run on a given hardware platform. With basic inference working, further improvements in speed, memory use and communication between servers will affect efficiency and service costs at scale.

▮ SIGNALOpen weights enable independent deployment, while GPU software adaptation broadens the choice of platforms on which the model can run.