Overview

Kimi K3 is Moonshot AI's most capable model with 2.8 trillion parameters, built on Kimi Delta Attention (a hybrid linear attention mechanism) and Attention Residuals. It offers native visual understanding, a 1M-token context window, and frontier intelligence for software engineering, knowledge work, and deep reasoning — served via Novita AI serverless inference. This circuit provides an OpenAI-compatible chat completion endpoint backed by the moonshotai/kimi-k3 model.

Inputs

1. Messages (messages)

• Description: A list of messages comprising the conversation so far.
• Type: array
• Required: Yes

2. Max Tokens (max_tokens)

• Description: The maximum number of tokens to generate in the completion. Kimi K3 supports up to 1048576.
• Type: number
• Default Value: 128000
• Required: No

3. Reasoning Effort (reasoning_effort)

• Description: Controls reasoning depth. Reasoning is always on for Kimi K3 and cannot be disabled, only adjusted via this parameter.
• Type: string
• Default Value: max
• Allowed Values: low, high, max
• Required: No

4. Tools (tools)

• Description: A list of tools the model may call. Supports function-type tools with JSON Schema definitions.
• Type: array
• Default Value: []
• Required: No

5. Tool Choice (tool_choice)

• Description: Controls how the model selects a tool. Supported: auto, none, required.
• Type: string
• Default Value: auto
• Required: No

Outputs

1. Chat Completion Output (chat_completion_output)

• Description: The generated chat completion text — the model's final answer.
• Value: Accumulated from choices[0].delta.content stream chunks.

2. Reasoning Output (reasoning_output)

• Description: The chain-of-thought reasoning content — the model's thinking process, separate from the final answer.
• Value: Accumulated from choices[0].delta.reasoning_content stream chunks.

3. Chat Completion Tool Calls (chat_completion_tool_calls)

• Description: Tool calls made by the model (if any).
• Value: Accumulated from choices[0].delta.tool_calls stream chunks.

Try It

You can try this circuit directly in the Playground — no setup required.

Notes

1. Reasoning Always On: Reasoning is always on for Kimi K3 and cannot be disabled. Use reasoning_effort to control depth — max for complex reasoning tasks, low for simple queries.
2. Fixed Parameters: Kimi K3 fixes temperature at 1.0, top_p at 0.95, and presence_penalty/frequency_penalty at 0. These are not exposed as circuit inputs — do not pass them explicitly.
3. Multimodal Input: Kimi K3 supports text, image, and video input. Send image/video as content blocks in the messages array.
4. Long Context: 1M token context window with up to 1M output tokens, suitable for long-document processing and complex engineering tasks.
5. Function Calling: Supports function calling and structured output via the tools and tool_choice inputs. tool_choice: "required" is supported.
6. Preserved Thinking: K3 always uses Preserved Thinking — reasoning content is emitted in reasoning_content and should be passed back in multi-turn conversations.

Enjoy .....!