Moonshot AI releases Kimi K3.1 Pro model weights: what's new

Moonshot AI, one of China's most prominent AI laboratories, has officially released the full weights for its Kimi K3.1 Pro model on Hugging Face as of November 4. Previously restricted to API and proprietary application access, the model is now available for broader research and deployment, meeting the company's commitment to a pre-November release deadline.

• The model is based on a 2.8 trillion parameter MoE (Mixture of Experts) architecture, with 104 billion active parameters per token.
• It supports a 1 million token context window and is capable of processing both text and image inputs.
• Distributed in a quantized format (MXFP4 for weights, MXFP8 for activations), the model takes up 1.56 TB of space.
• The training data was filtered to mitigate potential cybersecurity threats, slightly reducing the model's exploitation capabilities while maintaining peak performance.

Hardware and Deployment Challenges

Running Kimi K3.1 Pro is a resource-intensive endeavor. Given its 1.56 TB size, it does not fit into a single node with 8x B200 GPUs. A practical deployment requires at least eight B300 accelerators or two H200 nodes. On Hopper architecture, the lack of native FP4 hardware support might lead to inference performance bottlenecks.

Furthermore, the 1 million token context requires massive overhead for KV-cache. Consequently, self-hosting this model is reserved for organizations with high-end cluster capabilities rather than standard enterprise servers.

Performance and Competitive Positioning

Moonshot AI made a deliberate trade-off by sanitizing data related to cyberattacks, which reduced its vulnerability exploitation success rate from 90.5% to 81%. Despite this, the model consistently outperforms DeepSeek V4.1 Flash on the independent Vals Index and nears the performance levels of Claude Opus 5.5 in agentic coding tasks.

The model appears to be highly optimized for coding and agent-based workflows, though it may lag behind its predecessors in certain specialized legal or tax-related tasks, suggesting a strategic shift in optimization focus.

Significance for the Tech Ecosystem

The release of Kimi K3.1 Pro narrows the gap between open-weight models and top-tier proprietary systems. For businesses, this release is significant as it allows for secure, local deployment of advanced AI agents, enabling data privacy without sacrificing intelligence. This competitive push drives the industry toward more accessible yet powerful AI architectures.