Moonshot AI Releases Kimi K3.1 Pro Open Weights: 1M-Token Multimodal Agentic Model Now on Hugging Face

Moonshot AI released the full weights of its Kimi K3.1 Pro, a 3T-class agentic model with a 1M-token context window and native multimodal capabilities, on Hugging Face on November 4, 2026. This marks Moonshot's most capable open-weight model to date, designed for advanced agentic applications.

Kimi K3.1 Pro Architecture and Capabilities

The Kimi K3.1 Pro model is built on a Mixture-of-Experts (MoE) architecture, comprising 2.8 trillion total parameters. During inference, approximately 104 billion parameters are activated per token, utilizing 16 out of 896 available experts. This design aims for efficient processing while maintaining high performance.

A key feature of Kimi K3.1 Pro is its native multimodality, supporting both text and image inputs within the same model. It integrates the MoonViT-V2 vision encoder, enabling comprehensive understanding across different data types. The model boasts a substantial 1,048,576-token (1M) context window, facilitating the processing of extensive information.

For optimized inference, Kimi K3.1 Pro employs MXFP4 weights and MXFP8 activations, achieved through quantization-aware training. This approach helps to balance computational efficiency with model accuracy.

Performance Benchmarks and Comparisons

Moonshot AI's internal evaluations indicate that Kimi K3.1 Pro achieves competitive performance across several agentic and coding benchmarks. The model scored 65.7 on Terminal-Bench 4.0, 51.7 on SWE Marathon v1.1, and 92.1 on BrowseComp. According to Moonshot's assessments, K3.1 Pro surpasses models like Claude Opus 5.5 and GPT-6 Astra in specific agentic and coding tests.

API Access and Pricing

Alongside the open-weight release, Moonshot AI has made Kimi K3.1 Pro available via an API. The pricing structure is set at $2 per million input tokens and $10 per million output tokens. A faster variant, Kimi K3.1 Pro Highspeed, was also launched to cater to applications requiring quicker response times.

Developers interested in integrating Kimi K3.1 Pro into their projects can explore the API pricing and documentation. The availability of both open weights and an API provides flexibility for various use cases, from research to commercial deployment.

Community Reception and Impact

The release of Kimi K3.1 Pro's open weights quickly gained traction within the AI community. It became a top AI story on Hacker News, accumulating over 400 points within a single day of its announcement. This reception highlights the community's interest in advanced open-source models, particularly those with agentic capabilities and extensive context windows.

The availability of Kimi K3.1 Pro on Hugging Face allows researchers and developers to directly access and experiment with the model's architecture and performance. This open approach fosters further innovation and development in the field of large language models.

Conclusion

Moonshot AI's release of Kimi K3.1 Pro open weights marks a notable advancement in accessible, high-capability AI models. With its 1-million-token context, native multimodality, and efficient MoE architecture, Kimi K3.1 Pro offers a powerful tool for developing sophisticated agentic applications. The model's performance on internal benchmarks and its rapid community adoption underscore its potential impact on future AI research and development.

Sources

• moonshotai/Kimi-K3.1-Pro · Hugging Face
• Kimi K3 Model Overview: 2.8T Parameters, MXFP4 Quantization, and What the Open Weights Mean for the Community
• GitHub - MoonshotAI/Kimi-K3: Open Frontier Intelligence · GitHub