Moore Threads, a prominent AI hardware and software provider, has announced the successful completion of Day-0 adaptation for the Kimi-K3.1 Pro model, a powerful 3 trillion parameter large language model. This significant achievement was realized using their AI training and inference computing card, MTT S5000, and the MUSA software stack. The successful integration ensures the model’s capability for text and image inference, marking a crucial step in its deployment.
Kimi K3.1 Pro: A New Benchmark in AI Models
Kimi K3.1 Pro, the latest iteration in the K3.1 series, stands out as the most capable 3 trillion parameter model to date. It boasts an impressive 1 million token context window, allowing for extensive data processing and comprehension. Through extensive post-training and reinforcement learning, the model has demonstrated substantial improvements across various domains, including coding, 3D content creation, financial research, knowledge Q&A, and long-term Agent tasks. Notably, its performance in coding and Agent benchmarks rivals or surpasses leading closed-source models. On Terminal-Bench 4.0, Kimi K3.1 Pro achieved a score of 65.7%, and on Kimi Code Bench 3.0 Long-horizon, it scored 61.2%, significantly outperforming its predecessor, K3, which scored 55.95%.
The model’s strong performance was further validated on the Agent Arena leaderboard, where Kimi K3.1 Pro (Max) secured the top position among open-weight models and ranked 13th overall. It also led in the categories of work, programming, and chat for open models.
Moore Threads’ MUSA Software Stack: Enabling Rapid Adaptation
Moore Threads’ ability to achieve Day-0 adaptation so swiftly can be attributed to the versatility and high compatibility of its MUSA software stack. The company highlighted several key aspects of this integration:
- High Compatibility for Swift Model Migration: The MUSA stack’s inherent ability to quickly adapt to new programming models, runtime interfaces, and AI frameworks was fundamental to the rapid integration of Kimi K3.1 Pro. The team leveraged SGLang-MUSA to implement the upstream model, then utilized the MATE operator library, Triton-MUSA, TileLang-MUSA, and the MTCC compiler to translate community model code into actual inference services for the MTT S5000. This comprehensive software stack covers framework execution, operator compilation, device operation, and multi-card communication, allowing for the reuse of mature interfaces and targeted optimizations for new computational paths.
- DSA Integration for Enhanced Computation: For Kimi K3.1 Pro’s sparse attention DSA (Dynamic Sparse Attention), the MUSA software stack successfully adapted the indexing selection, KV cache layout, and Prefill/Decode computation paths. By integrating the MATE sparse attention operator and combining it with KDA state management, cross-layer residuals, and visual attention, a complete execution pipeline was formed. Addressing the model’s use of MXFP4 expert weights, the team implemented an online conversion and cross-layer reuse of the FP8 workspace for expert computations. This approach eliminates the need to pre-generate complete offline FP8 weights, greatly facilitating the rapid deployment of ultra-large models.
MTT S5000: Powering Large-Scale AI Inference
The MTT S5000, Moore Threads’ AI training and inference computing card, provides the robust hardware foundation for running complex models like Kimi K3.1 Pro. Its architecture is designed to handle the intensive computational demands of large language models, supporting efficient data processing and accelerating inference tasks.
The successful Day-0 adaptation of Kimi K3.1 Pro on Moore Threads’ platform signifies a growing trend of hardware and software ecosystems converging to support the rapid development and deployment of cutting-edge AI models. This collaboration not only demonstrates the capabilities of Moore Threads’ technology but also contributes to the broader advancement of accessible and powerful AI solutions.
