Kimi K3.1 Pro by Moonshot (1M context)

Overview

Kimi K3.1 Pro is Moonshot’s open-weight large language model in the K3.1 series. It takes text plus optional images, then produces a text response that can include structured JSON or tool calls. It’s post-trained from Kimi K3 with large-scale reinforcement learning for long, multi-step work. The 1,048,576-token context window makes it useful for large codebases, long documents, and agent loops that must keep state.

What you can build

• Code review and refactor runs across large repositories, with multi-step plans and tool-assisted execution
• Long-document Q&A and research briefs over large reports, specs, or logs
• Chart, screenshot, and UI analysis from images, with written findings and action items
• Agent workflows that call functions for search, retrieval, calculations, or business logic
• JSON-first extraction pipelines for invoices, tables, checklists, and compliance summaries

Inputs

• A single instruction written in plain text.
• Or a multi-turn chat transcript as an ordered list of user and assistant messages.
• Up to 10 optional reference images per request. Use JPEG, PNG, or WebP files up to 10 MB each.
• Optional system-level instructions that set role, tone, and rules for the whole session.
• Optional user and session identifiers (string or number) to keep conversation history consistent across turns.
• Optional “thinking” intensity. Higher settings spend more tokens on reasoning before the final answer.
• Optional web search toggle. When enabled, the model may search the web during the run.
• Optional tool definitions for function calling. Provide function specs with JSON Schema-style parameter definitions.
• Optional tool-choice behavior. You can let the model decide, forbid tools, or require tool use.
• Optional output token budget. Set a cap from 16 up to 1,048,576 tokens.
• Optional continuation inputs for tool-result follow-ups. Provide the prior run token plus the tool results so the model can continue the same turn.

Outputs

Kimi K3.1 Pro returns a text response meant to be shown directly to a user or passed to an agent step.

If you enable tools, the model can return one or more function-call requests. Each request includes a tool name plus JSON arguments that match your tool’s schema.

If you request structured output, the response can be constrained to valid JSON (including schema-constrained JSON when supported by the route).

Limitations

• Text output only. It can analyze images, but it doesn’t generate image files.
• Video inputs aren’t supported on this model route.
• The full context window is 1,048,576 tokens, including your input, tool text, and the model’s reasoning and answer.
• Sampling controls may be fixed on some endpoints. Requests that try to override them can fail.
• Long runs can produce very large outputs. Set an output cap when you don’t need exhaustive reasoning.
• Low-quality inputs reduce accuracy. Blurry screenshots, tiny text, heavy compression, and skewed scans can break OCR and visual grounding.

Safety & compliance

Follow applicable laws and your organization’s policies when you use the model.

Don’t use it to create malware, exploit instructions, or instructions for wrongdoing. Treat any security-related output as untrusted and review it.

If you enable web search, verify any claims against primary sources before you act on them.

If you include personal data, keep it minimal. Remove sensitive identifiers when possible, especially in logs and screenshots.