Moonshot AI has introduced Kimi K3.1 Pro, the first model in its new K3.1 series and what the company describes as its most capable open-weight model to date. The system features a one-million-token context window and is positioned to compete with leading proprietary models across coding, 3D creation, and financial research tasks.
The release marks a significant step in open-weight AI development. Unlike closed systems where model weights remain proprietary, open-weight releases allow researchers and developers to download and run models on their own infrastructure. Moonshot AI plans to release the full K3.1 Pro weights by November 4, 2026.
Performance across specialized domains
In standardized evaluations, K3.1 Pro demonstrates considerable capability in software engineering tasks. On Terminal-Bench 4.0, which tests coding agents on realistic terminal-based programming tasks, the model scored 65.7 percent - ahead of GPT-6 Astra at 58.2 percent and just behind Claude Opus 5.5 at 66.4 percent. The company also reports strong results on BenchCAD, a test of 3D reconstruction from multi-view images, where K3.1 Pro achieved 92.7 percent with tools enabled.
Financial analysis represents another focus area. The model leads on OfficeQA Pro v2, a benchmark for complex fiscal data analysis across documents, with a score of 64.4. Moonshot AI reports that hallucination rates dropped substantially from the previous K3 version - from 23.4 percent to 8.3 percent on internal factual accuracy tests.
The company is also releasing Kimi K3.1 Pro Highspeed, which offers two to three times the response speed of the base model. Both versions are available through the Kimi website, desktop applications, and API, with enterprise controls that default to off for organizational access.
This development matters because it continues the trend of narrowing capability gaps between open-weight and proprietary AI systems while maintaining lower deployment costs. For organizations concerned about data privacy or requiring offline operation, open-weight models provide alternatives to API-dependent services. The release also includes deliberate limitations on offensive cybersecurity capabilities through training data removal - an approach that may inform how future powerful models balance openness with safety considerations.