Kimi K3.1 Pro

Supported Functionality

Item	Specification
Input	Text, image, video
Output	Text
Context	1,048,576 tokens
Max Output	131,072 tokens by default; up to 1,048,576 tokens
Vision	✓ Supported
Function Calling	✓ Supported

Description

Kimi K3.1 Pro is the first model in Moonshot AI's K3.1 series, released on October 23, 2026 under the API model code kimi-k3.1-pro. It is a heavily post-trained, 3T-class open-weight flagship model with a 1,048,576-token context window, text, image, and video input, and text output. Moonshot plans to release the full model weights by November 4, 2026.

The model improves coding, long-running agents, 3D modeling and design, financial research, and professional knowledge work. Thinking is always enabled and the top-level reasoning_effort accepts low, high, or the default max. It also supports separate streaming deltas for reasoning and answers, strict JSON Schema, Partial Mode, custom tools, tool_choice, dynamic tool loading, and automatic context caching. The separate kimi-k3.1-pro-highspeed variant is officially described as approximately two to three times faster.

Key Capabilities

• Real-Repository Development: Covers quick fixes, multi-file changes, complete project builds, debugging, and iterative repair.
• Long-Running Agent Execution: Analyzes results, diagnoses problems, calls tools, and continues refining work across multiple stages.
• 3D Modeling and Spatial Reasoning: Infers shape, proportion, and spatial relationships from text, sketches, or images and writes scripts for Blender and CAD tools.
• Financial Research and Modeling: Handles real documents, cross-source verification, numerical reasoning, cross-sheet formulas, and end-to-end investment workflows.
• Image and Video Understanding: Analyzes local images and uploaded video for visual questions, summaries, and cross-media tasks.
• Structured and Prefix-Constrained Generation: Uses strict JSON Schema for final answers and Partial Mode to continue from a supplied prefix.
• Agentic Tool Orchestration: Supports custom functions, required tool choice, and tool definitions loaded dynamically within message history.

Technical Strengths

Feature	Benefit
1,048,576-Token Context	Fits large repositories, long videos, financial documents, and extended agent history within one task.
Up to 1,048,576 Output Tokens	Defaults to 131,072 and can be raised for large code or long professional deliverables.
Three Always-On Reasoning Levels	low, high, and default max adjust depth and responsiveness, while thinking cannot be disabled.
Native Image and Video Input	Combines text, images, and video for interface, spatial, motion, and multimedia analysis.
Strict Structured Output	json_schema with strict: true improves machine parsing and downstream reliability.
Automatic Context Caching	Eligible prefixes are cached automatically with five-minute and one-hour TTL options.

Frequently Asked Questions

How should I choose low, high, or max reasoning in Kimi K3.1 Pro?

Use low for simple analysis, fast tool calls, and latency-sensitive work, high for complex coding and professional analysis, and the default max for the hardest long-running agents, mathematics, or large engineering tasks. Compare correctness, first-pass completion, tool loops, output length, and response time on representative workloads.

Can thinking be disabled in Kimi K3.1 Pro?

No. Thinking is always enabled; if reasoning takes too long, set reasoning_effort to low rather than sending an unsupported disable setting.

What is the difference between the one-million-token context and maximum output?

The 1,048,576-token context holds input and conversation history, while max_completion_tokens defaults to 131,072 and can be set as high as 1,048,576. Effective room still depends on the complete request and service limits, so very long output needs streaming, reconnection, and completeness checks.

How do I call Kimi K3.1 Pro on LinkModel?

Kimi K3.1 Pro is available on LinkModel. Use the public model ID, request schema, context limits, reasoning levels, and multimodal fields shown on LinkModel's live model page, beginning with small text and streaming tests before adding images, video, and tools.

How should I validate the Kimi K3.1 Pro integration on LinkModel?

Test all three reasoning levels, long context, very long output, images, video, strict JSON, Partial Mode, and multi-turn tools, including timeouts, refusals, tool failures, and retries. Do not assume Moonshot-native reasoning_content, file URIs, cache TTL, dynamic tools, or fixed sampling behavior pass through LinkModel unchanged.