Chinese company Moonshot AI unveiled Kimi K3.1 Pro on Oct. 23, 2026, saying the updated artificial intelligence model improves programming, financial analysis and 3D modeling. The release is intended to keep Moonshot competitive in advanced AI after its previous model, Kimi K3, drew industry attention in the summer of 2026.

K3.1 Pro has a 1 million-token context window and a mixture-of-experts architecture with 2.8 trillion total parameters, of which 104 billion are active. The multimodal model is already available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API, while Moonshot expects to publish its full weights by Nov. 4, 2026.

Moonshot said it used significantly more post-training compute for K3.1 Pro than for K3. The model can work on complete software projects, including games built with Unreal Engine, Unity and Godot, as well as interactive webpages and 3D scenes.

Programming, 3D and financial benchmarks

On Terminal-Bench 4.0, which evaluates AI agents on realistic terminal-based tasks, K3.1 Pro scored 65.7%. GPT-6 Astra scored 58.2%, while Claude Opus 5.5 scored 66.4%. On Moonshot’s internal Kimi Code Bench 3.0 Long-horizon test, the new model scored 61.2%, compared with 55.95% for K3.

In 3D modeling, K3.1 Pro scored 92.7% with tools on BenchCAD, a test that evaluates the reconstruction of three-dimensional objects from images taken at different angles. Claude Opus 5.5 scored 96.2%, and GPT-6 Astra scored 95.9%.

Moonshot’s published results show that K3.1 Pro scored 64.4 points on OfficeQA Pro v2, which tests financial-document analysis, data reconciliation and multi-step calculations. Its score on the company’s internal Financial Modeling test rose by 7.2 points from K3, while its Business Operation performance was about 4.2 times higher.

Hallucination and cybersecurity tests

On Moonshot’s internal Kimi Hallu Bench Hard test, the share of hallucinated statements declined from 23.4% for K3 to 8.3% for K3.1 Pro. The company cautioned that the benchmark deliberately uses difficult tasks designed to provoke errors, meaning the results do not represent the model’s overall error rate in ordinary use.

Moonshot said it removed cybersecurity-related data from two training stages while preparing the open version. On an internal vulnerability-discovery and exploitation test, the final K3.1 Pro completed 81% of tasks in pass@3 mode, compared with 90.5% for a control version without the restrictions. The company said the measures reduce the risk of misuse while preserving the model’s usefulness for general programming.

Moonshot also offers K3.1 Pro Highspeed, which the company says operates two to three times faster than the standard version.

Funding and regulatory backdrop

Bloomberg reported that Moonshot completed a private funding round at a valuation of about $50 billion and is considering a Hong Kong Stock Exchange listing in the first quarter of 2027. Its annual recurring revenue run rate exceeded $1 billion in August, up from $300 million in June, and the company aims to raise that figure to $2 billion by the end of 2026.

Moonshot has also faced allegations that it used outputs from larger models to train its systems. Anthropic alleged in a September report that Moonshot had violated its rules by routing user prompts through Claude. Moonshot did not respond publicly to the allegations.

Chinese regulators have separately proposed requiring registration before China-developed AI technologies are introduced in overseas markets. State media reported that DeepSeek and MiniMax had already received approval to deploy their models internationally under the draft rules.

Source: Incrypted