Kimi K3.1 Pro: Moonshot's 1M-Token Model Targets Coding and AI Agents

Kimi K3.1 Pro: Moonshot's 1M-Token Model Targets Coding and AI Agents#

Moonshot AI has introduced Kimi K3.1 Pro, a large-scale reasoning model designed for long-horizon coding, 3D creation, financial research, and complex agentic workflows. The model offers a 1-million-token context window and is a post-trained version of Kimi K3, developed using substantially more post-training compute.

The release also introduces Kimi K3.1 Pro Highspeed, a faster variant that Moonshot says delivers two to three times the speed of the standard model.

As of November 1, 2026, both variants are available through Moonshot's product ecosystem and API. However, the full K3.1 Pro model weights have not yet been publicly released. Moonshot has scheduled their release by November 4, 2026.

The distinction matters for developers: Kimi K3.1 Pro can already be evaluated through hosted services, but teams planning to run it on their own infrastructure must wait for the weights and review the accompanying license.

What Did Moonshot Announce?#

Kimi K3.1 Pro is the first model in Moonshot's K3.1 series. Rather than introducing an entirely new base model, Moonshot describes it as a substantially post-trained version of Kimi K3.

The development effort focuses on improving reasoning and the ability to complete tasks involving multiple steps, external tools, and extended execution.

Moonshot identifies four important application areas:

• Software engineering: Debugging, repository-level development, code generation, and long-running programming tasks.
• 3D creation: Parametric modeling, spatial reasoning, and CAD-oriented workflows.
• Financial research: Document analysis, financial modeling, and multi-stage investment workflows.
• Agentic knowledge work: Research and other tasks that require planning, tool execution, and iterative problem-solving.

The release is positioned around completing larger workflows rather than simply improving short-form conversational responses.

Source: Moonshot AI's Kimi K3.1 Pro announcement.

Kimi K3.1 Pro: Technical Specifications#

The published specifications describe a model in the 3-trillion-parameter class with a particularly large context window.

Moonshot describes the model as a 3T-class system. The 2.8T total and 104B active parameter figures are listed by Artificial Analysis.

These are different measurements. Total parameters describe the overall model size, while active parameters refer to the parameters activated during a particular inference computation.

Source: Artificial Analysis model comparison.

Understanding the Model Architecture#

An important part of the Kimi K3.1 Pro announcement is its relationship with the original Kimi K3 architecture.

In its earlier K3 announcement, Moonshot described an architecture incorporating:

• Kimi Delta Attention (KDA): An attention mechanism designed to improve information processing across long sequences.
• Attention Residuals (AttnRes): An architectural approach intended to improve information flow across model depth.
• Stable LatentMoE: A sparse Mixture-of-Experts approach that activates selected experts rather than processing every expert for every token.

The K3 architecture description identifies 16 activated experts from a pool of 896.

Moonshot's K3.1 announcement does not introduce a separate architecture specification. Instead, it describes K3.1 Pro as a post-trained K3 model.

Consequently, the earlier K3 architecture provides useful context for understanding the model family, but it should not be mistaken for a complete, independently published K3.1 Pro architecture report.

Source: Moonshot's Kimi K3 architecture announcement.

What Changed Technically in Kimi K3.1 Pro?#

The central development is the increased investment in post-training and reinforcement learning.

Moonshot says it used substantially larger post-training compute and optimized multiple parts of its training and inference infrastructure.

These include:

• Reinforcement-learning frameworks.
• Long-context inference.
• Training efficiency.
• Sandbox infrastructure.
• Rollout scheduling.
• Environment-quality evaluation.

The company describes a large-scale joint reinforcement-learning process across hundreds of downstream domains.

Why Reinforcement Learning Matters#

Pretraining gives a language model its general ability to understand and generate information. Post-training then improves how the model responds to instructions and handles specific tasks.

Reinforcement learning can further optimize model behavior using feedback from evaluated actions or outcomes.

For agentic systems, this is particularly relevant.

Consider an AI coding agent asked to fix a problem in a large software project. Completing the task may involve:

1. Understanding the repository structure.
2. Identifying the relevant source files.
3. Investigating the underlying problem.
4. Editing the implementation.
5. Running tests.
6. Analyzing failures.
7. Revising the implementation.
8. Verifying the final result.

A model that can reason effectively across this sequence has a practical advantage over one that only generates plausible code snippets.

Moonshot's K3.1 development strategy places considerable emphasis on these longer execution sequences.

However, reinforcement learning alone does not make a model a complete autonomous software engineer. The surrounding execution environment, available tools, permissions, testing infrastructure, and verification mechanisms remain essential.

Coding, 3D Creation, and Financial Research#

Moonshot highlights improvements across three major technical areas.

1. Long-Horizon Software Engineering#

K3.1 Pro is designed to handle larger coding tasks that require sustained reasoning, tool use, and iterative execution.

The model's coding capabilities extend beyond generating individual functions. Moonshot describes applications involving complete projects, debugging, game development, and interactive web experiences.

The company also reports improvements in a long-horizon coding evaluation called Kimi Code Bench 3.0.

K3.1 Pro scored 61.2%, compared with 55.95% for Kimi K3.

This is a Moonshot-reported benchmark, not an independent measurement.

For developers, the more important question is whether the model can consistently manage a software task from initial investigation to verified completion.

That requires evaluation across real repositories, automated tests, tool reliability, and the number of human interventions required.

2. 3D Modeling and CAD Workflows#

Another focus is generating and refining 3D assets.

K3.1 Pro can work with text descriptions, sketches, and reference images to infer spatial relationships and generate assets using tools such as Blender and CAD software.

Moonshot reports a score of 92.7% on BenchCAD with tools.

The benchmark evaluates 3D reconstruction through CAD code generation.

The company also reports:

These are company-reported comparison results. They should not be interpreted as a universal ranking across every 3D modeling workflow.

The practical significance is that language models can increasingly participate in structured design workflows, where they generate scripts, inspect intermediate output, and refine geometry through external tools.

3. Financial Research and Document Analysis#

Moonshot also emphasizes financial research and complex document processing.

K3.1 Pro reportedly achieved 64.4 on OfficeQA Pro v2, an evaluation involving financial documents, retrieval, cross-document verification, and numerical reasoning.

The company reports improvements in its internal financial modeling and research evaluations as well.

This capability could be relevant to applications such as:

• Financial document processing.
• Automated spreadsheet analysis.
• Investment research assistance.
• Business performance reporting.
• Enterprise knowledge extraction.

However, financial applications require additional safeguards.

An LLM-generated financial conclusion should be traceable to its source documents, numerical calculations should be independently checked, and high-impact decisions should retain human review.

Kimi K3.1 Pro Benchmark Results#

Benchmark results provide useful evidence about model capabilities, but their interpretation depends on evaluation methodology.

Moonshot publishes its own comparison results, while Artificial Analysis provides independent evaluations using its own testing frameworks.

These sources should be considered separately.

Moonshot-reported results#

Moonshot reports that the model uses the Claude Code harness for Terminal-Bench 4.0.

Its published benchmark notes also specify maximum reasoning effort, temperature 1.0, and top-p 0.95 unless otherwise stated.

These settings matter because reasoning effort, inference configuration, and agent harnesses can influence benchmark outcomes.

Source: Moonshot benchmark and methodology notes.

Independent evaluations#

Artificial Analysis provides a separate view of model performance.

Its reported results include:

These results suggest a significant improvement in coding-agent and agentic SaaS workflow evaluations.

Artificial Analysis also publishes an Intelligence Index. Its October 23 launch report initially listed K3.1 Pro at 55. Its currently displayed model comparison uses a newer v5.0 index and lists a score of 46.

These are different index snapshots and should not be presented as a direct performance decline. The current composite index and individual benchmark results should be checked separately.

Sources:

• Artificial Analysis launch evaluation
• Artificial Analysis current model comparison.

An independent data-analysis test#

Plotly also evaluated K3.1 Pro using its own data analytics examination.

The model passed 42 out of 43 questions in that evaluation.

Its test included conventional analytics, ambiguous requests, data-quality checks, and hidden-pattern discovery.

This is useful evidence for evaluating tool-assisted analytical work, but it is not a general-purpose benchmark covering every model capability. The evaluation uses a private synthetic dataset.

Source: Plotly's Kimi K3.1 Pro evaluation.

What Developers Can Build With Kimi K3.1 Pro#

The model's most relevant applications involve workflows where an LLM must coordinate multiple reasoning and execution steps.

AI coding assistants#

K3.1 Pro can serve as the reasoning component of a coding assistant that navigates repositories, investigates bugs, modifies source files, and runs tests.

A production implementation would typically require:

• Repository indexing and file access.
• A controlled terminal or execution sandbox.
• Permission management.
• Automated testing.
• Tool-call monitoring.
• Human approval for sensitive operations.

The model is only one part of this architecture.

Enterprise document intelligence#

The 1M-token context window and document-analysis capabilities could support applications involving large technical documents, financial reports, and internal knowledge bases.

Potential workflows include document comparison, information extraction, research synthesis, and evidence-based question answering.

However, a large context window does not eliminate the need for retrieval, access control, document versioning, and source-grounded evaluation.

For enterprise systems, a RAG architecture may still be preferable when documents are frequently updated or users have different data-access permissions.

AI research assistants#

Research agents can combine model reasoning with search, retrieval, structured extraction, and calculation tools.

K3.1 Pro's focus on multi-step reasoning makes it a candidate for systems that investigate technical subjects, analyze information from multiple sources, and produce structured reports.

Such applications should retain source citations and independently verify important claims.

Automated design and 3D workflows#

Developers can integrate the model with graphics and CAD environments to create systems that generate scripts, manipulate assets, and iterate on 3D designs.

The model's performance will depend on the external application interface, available tools, and quality of the feedback provided during each iteration.

API Access, Pricing, and Availability#

Kimi K3.1 Pro is available through Moonshot's product ecosystem and API.

Developers can select the model identifiers:

• kimi-k3.1-pro
• kimi-k3.1-pro-highspeed

Moonshot's API platform currently lists the following pricing.

Highspeed is positioned as the faster option for applications that need more responsive interactions. Moonshot claims approximately two to three times the speed of the standard variant.

The higher token prices mean developers should evaluate the total cost of completed tasks, not just the input price.

For example, an agent that performs many reasoning steps, executes tools repeatedly, and generates lengthy responses may consume substantially more tokens than a conventional conversational application.

Where can developers access the model?#

Moonshot lists availability through:

• Kimi.com and the Kimi mobile applications.
• Kimi Work.
• Kimi Code.
• Kimi API.
• Kimi Business and supported enterprise deployments.

Moonshot also states that Zero Data Retention is available for API customers.

Enterprise developers should confirm the relevant data-handling terms, account configuration, and contractual conditions before using confidential business information.

Official pricing and access: Kimi API Platform.

Open Weights: What Developers Should Know#

One of the most important aspects of the announcement is the planned release of the full model weights.

Moonshot says the Kimi K3.1 Pro weights will be released by November 4, 2026.

As of November 1, the weights are still pending.

This means developers currently have hosted access, but cannot yet treat the K3.1 Pro release as an available self-hosted model.

A public weight release could give developers more control over deployment, inference infrastructure, customization, and data-processing environments.

However, downloadable weights do not automatically mean unrestricted open-source use.

The actual license must be reviewed before commercial deployment, redistribution, fine-tuning, or providing model-as-a-service offerings.

Developers should also evaluate hardware requirements, quantization support, inference frameworks, memory consumption, and serving throughput when the release becomes available.

Cybersecurity and Safety Considerations#

K3.1 Pro's development also highlights an important challenge for increasingly capable open-weight models.

Models that can write code and operate tools may be useful for defensive security research, but the same capabilities can potentially support vulnerability exploitation.

Moonshot says it removed cyber-related data from parts of the training corpus and filtered cyber-related tasks from its reinforcement-learning environments for the planned open-weight release.

The company reports that its internal xdev-bench-hard evaluation produced an 81.0% result for the risk-reduced K3.1 Pro, compared with 90.5% for a similar checkpoint without those risk-reduction measures.

These are company-reported results rather than independently verified security measurements.

The distinction is important because safeguards applied to hosted models are easier to update centrally than restrictions embedded in a distributed model's training and weights.

Once weights are publicly distributed, Moonshot has less control over how downstream users deploy or modify them.

For organizations planning to deploy K3.1 Pro, security evaluations should include:

• Testing model behavior with untrusted inputs.
• Restricting tool permissions and execution environments.
• Logging and auditing agent actions.
• Separating development, testing, and production credentials.
• Evaluating both defensive and potentially harmful security capabilities.

Source: Reuters coverage of Moonshot's cybersecurity restrictions.

Limitations and Engineering Trade-offs#

Despite its capabilities, Kimi K3.1 Pro should not be treated as a universally suitable model.

Several practical considerations deserve attention.

Inference latency: Longer reasoning and multi-step agent execution can increase response times. Artificial Analysis reports approximately 31 output tokens per second and around 68 seconds to the first answer token in its current standard-model evaluation. These figures depend on the specific testing configuration and should not be interpreted as guaranteed API performance.

Inference economics: The standard model is priced at $2 per million input tokens and $10 per million output tokens. Large reasoning traces and repeated tool calls can increase the overall cost of an application.

Context management: A 1-million-token context window provides room for large inputs, but it does not guarantee that every relevant detail will be retrieved or used correctly. Retrieval, summarization, and evaluation remain important.

Self-hosting uncertainty: The K3.1 Pro weights and their license terms are still pending as of November 1, 2026. Teams cannot yet make a fully informed self-hosting or licensing decision.

Benchmark limitations: Company-reported results, independent benchmark evaluations, and application-specific performance are different measurements. Real deployments should use representative workloads.

These trade-offs make application-level testing essential before choosing K3.1 Pro for a production system.

What Kimi K3.1 Pro Means for AI Engineering#

Kimi K3.1 Pro reflects a broader direction in foundation-model development: the industry is increasingly focusing on models that can execute longer tasks instead of simply producing better individual responses.

Its improvements are aimed at software engineering, research, numerical analysis, and workflows involving multiple tools.

For developers, the relevant question is not whether a model has trillions of parameters or a particularly large context window. It is whether the model can complete a specific workload reliably, at an acceptable cost, with manageable latency and appropriate security controls.

K3.1 Pro is worth evaluating for long-horizon coding agents, document-heavy enterprise applications, and tool-assisted analytical systems.

Its eventual open-weight release could make it more attractive to organizations seeking greater infrastructure control. That decision, however, should wait for the actual weights, license, deployment requirements, and independent testing results.

The November 4 weight-release milestone will be an important next step in determining how accessible the model becomes beyond Moonshot's hosted ecosystem.