Kimi K3.1 Pro Beats GPT-6 Astra in Terminal-Bench With 65.7% Score and Reaches 92.7% in Tool-Assisted 3D Modelling Test

Kimi AI has introduced Kimi K3.1 Pro, a new artificial intelligence model designed for advanced coding, 3D modelling, financial research and complex multi-step tasks. Developed by Moonshot AI, the model is the first release in the K3.1 series and is described by the company as its most capable 3-trillion-parameter-class open-weight model to date. It features a 1-million-token context window and improvements in reasoning, software development and agentic workflows.

The company has also launched Kimi K3.1 Pro Highspeed, a faster version designed to deliver approximately two to three times the speed of the standard model. Both models are available through Kimi.ai, Kimi Work, Kimi Code and the Kimi API. Moonshot AI said the full model weights are scheduled for release by November 4, 2026.

Kimi K3.1 Pro Features 3-Trillion-Class Architecture

Kimi K3.1 Pro uses a mixture-of-experts (MoE) architecture with approximately 2.8 trillion total parameters and 104 billion active parameters, according to the company’s announcement. Its 1-million-token context window allows it to process large volumes of information within a single context, supporting long documents, extensive codebases and complex research tasks.

The model is a post-trained version of Kimi K3, developed using additional post-training computation and large-scale reinforcement learning. Moonshot AI said it optimised its reinforcement learning framework, long-context inference, training efficiency and execution infrastructure to improve performance across a broad range of tasks.

The company reported that K3.1 Pro matches or approaches leading proprietary AI models across several coding, agentic and knowledge-work benchmarks. However, the published results are based on specific evaluation settings and do not necessarily indicate superior performance across every real-world task.

Kimi K3.1 Pro Benchmark Results Show Coding Improvements

Kimi K3.1 Pro recorded a score of 65.7% on Terminal-Bench 4.0, a benchmark that evaluates AI coding agents on practical terminal-based software engineering tasks. The score was higher than the reported 58.2% for GPT-6 Astra and slightly below Claude Opus 5.5’s 66.4%.

On Kimi Code Bench 3.0 Long-horizon, an internal benchmark designed to assess multi-step coding, tool usage and end-to-end problem-solving capabilities, K3.1 Pro scored 61.2%, compared with 55.95% for its predecessor, K3.

The model is designed to support a range of software development activities, from debugging existing codebases to building complete applications. It can work with game development engines such as Unreal Engine, Unity and Godot, as well as web technologies and 3D generation tools.

Moonshot AI also highlighted improvements in creating interactive web experiences, including Three.js scenes, animated interfaces and visual assets. These capabilities are intended to help developers complete larger projects that require planning, coding, testing and iterative refinement.

Kimi K3.1 Pro Targets Advanced 3D Modelling and Blender Workflows

Three-dimensional modelling is another major focus of the new model. Kimi K3.1 Pro is designed to interpret text descriptions, sketches and reference images to generate 3D assets while reasoning about geometry, proportions and spatial relationships.

On the BenchCAD benchmark, which evaluates 3D reconstruction through CAD code generation, K3.1 Pro achieved a score of 92.7% with tools. This compared with 96.2% for Claude Opus 5.5 and 95.9% for GPT-6 Astra in the company’s reported results.

The model can also generate scripts for Blender and CAD software to create or modify assets, adjust geometry, change materials and refine scene layouts. These capabilities could support workflows involving mechanical components, architectural structures, furniture, product visualisation and game assets.

Moonshot AI demonstrated the model’s ability to construct a large fantasy environment in Unreal Engine 5.8 featuring a procedurally modelled dragon created in Blender. The project included a 36-bay colonnade, a large-scale environment and rendering technologies such as Nanite and Lumen.

The company has also introduced 3D Make, a feature available through Kimi’s website and app. It allows users to preview and edit generated 3D models online, download models for 3D printing and export creations as videos or Live Photos.

Kimi K3.1 Pro Improves Financial Research and Modelling

Financial analysis and research are among the other areas targeted by Kimi K3.1 Pro. The model is designed to perform multi-step financial workflows involving document retrieval, cross-document verification, numerical analysis and spreadsheet modelling.

On OfficeQA Pro v2, a benchmark focused on complex financial document analysis, K3.1 Pro achieved a score of 64.4. It also recorded 58.6 on Frontier Finance, an evaluation focused on investment workflows and profitability analysis.

Moonshot AI reported improvements over Kimi K3 in its internal financial evaluations. The model’s financial modelling score increased from 50.4 to 57.6, while its financial research score rose from 58.8 to 70.4. Its business operations score increased from 8.9 to 37.

These results indicate improvements in the company’s internal tests of spreadsheet logic, evidence-based research and decision-making in simulated business environments. Independent validation would be needed to determine how consistently these gains translate into practical financial work.

Company Reports Lower Hallucination Rate in Financial Tasks

Moonshot AI said it made targeted improvements to the model’s factual accuracy, particularly for financial research and tasks requiring evidence-based answers.

On its internal Kimi Hallu Bench Hard evaluation, the claim-level hallucination rate declined from 23.4% for Kimi K3 to 8.3% for Kimi K3.1 Pro. The benchmark was designed to stress-test factual accuracy using challenging prompts and scenarios in which models are more likely to make errors.

The company cautioned that the evaluation does not represent the overall error rate during everyday use. The findings instead reflect performance under deliberately difficult testing conditions.

For financial applications, improvements in factual accuracy can be important because incorrect figures, unsupported assumptions or fabricated information may affect research conclusions and financial models. Users should still independently verify important calculations, market data and investment-related claims.

Moonshot AI Introduces Cybersecurity Safeguards

Moonshot AI also outlined measures intended to reduce potential cybersecurity misuse before releasing K3.1 Pro’s model weights.

The company said it identified and removed cybersecurity-related data from selected stages of training and filtered related reinforcement-learning tasks while retaining general-purpose coding capabilities.

On its internal xdev-bench-hard evaluation, K3.1 Pro achieved an 81.0% pass@3 score. The company reported that a version trained without the described risk-reduction measures achieved 90.5%, indicating that the interventions reduced performance on the tested vulnerability research and exploitation tasks.

Moonshot AI said its approach aims to make capable open-weight AI models available for legitimate uses, including defensive security research, while limiting the release of models that could substantially increase offensive cyber capabilities.

Kimi K3.1 Pro Pricing and Availability

Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed are available through the company’s product ecosystem and API platform. The API pricing published in the supplied launch information is as follows.

For the standard Kimi K3.1 Pro model, input tokens cost $2 per million tokens, while output tokens cost $10 per million tokens. Cache writes are priced at $2 per million tokens, and cache hits cost $0.20 per million tokens.

The Highspeed version costs $4 per million input tokens and $20 per million output tokens. Cache writes are priced at $4 per million tokens, while cache hits remain at $0.20 per million tokens.

The Highspeed model is intended for users who prioritise response speed, while the standard version offers a lower API price. Actual costs will depend on token usage and the applicable API billing rules.

Users can access K3.1 Pro through Kimi.ai and its mobile applications, Kimi Work for desktop, Kimi Code for software development and the Kimi API for integration into applications and automated workflows. Enterprise administrators can also enable access for their organisations.

Moonshot AI said the full model weights are expected to become available by November 4, 2026, which would allow developers to explore deployment and customisation options under the applicable release terms.

Kimi K3.1 Pro Expands AI Capabilities Across Development and Research

The launch of Kimi K3.1 Pro expands Moonshot AI’s efforts to develop AI systems capable of handling complex, multi-stage work rather than isolated prompts. Its combination of a 1-million-token context window, a large mixture-of-experts architecture and reported improvements in coding, 3D creation and financial research positions it for users working across software engineering and technical analysis.

The upcoming open-weight release could provide developers with additional flexibility to run and adapt the model, depending on its hardware requirements and licence conditions. Its practical adoption will depend on independently verified performance, inference costs, deployment requirements and the reliability of its outputs in real-world workflows.