Skip to main content
Last Updated: 2025-12-03
Status: Supported (OpenAI-Compatible)

​
1. Overview

Interleaved Thinking is an advanced reasoning framework that enables models to perform explicit reasoning steps between tool calls. Models with Interleaved Thinking can:
  • Reflect on the current environment and tool outputs
  • Decide the next action based on updated reasoning
  • Maintain a continuous reasoning chain across multiple tool invocations
  • Provide transparent, inspectable multi-step thinking through reasoning_details or reasoning_content
This capability transforms traditional function-calling into agent-level tool use, making complex workflows more accurate, reliable, and context-aware. Novita fully supports Interleaved Thinking for all models that natively expose reasoning streams (e.g., MiniMax-M2 and other OpenAI-compatible reasoning models).

​
2. Key Concepts

​
2.1 Interleaving

Instead of executing a single reasoning phase followed by a tool call, the model performs: 
Reason → Tool Call → Observe → Reason → Tool Call → ...
This allows the model to adjust its strategy dynamically based on previous tool outputs.

​
2.2 Reasoning Details (reasoning_details)

For some models, the content of the model’s thinking will be returned in the form of a separate structure: 
"reasoning_details": [
  {
    "type": "reasoning.text",
    "format": "openai-responses-v1",
    "text": "Model’s step-by-step reasoning..."
  }
]
For such models, Novita supports returning this field in both streaming and non-streaming modes.

​
2.3 Conversation Memory Requirement

To maintain reasoning continuity: You must append the model’s full response including reasoning_details, tool_calls, and content to subsequent messages. Failing to preserve the chain may result in:
  • Incorrect tool use
  • Lost reasoning context
  • Repeated or circular tool calls
  • Reduced reliability
This requirement mirrors OpenAI’s reasoning APIs.

​
3. API Behavior

​
3.1 Request Format

No changes are required on the user side. Interleaved Thinking works with the standard OpenAI-compatible Chat Completions API.

​
3.2 Response Format

The model may return the following fields:
  • reasoning_content: original thinking content
  • reasoning_details: structured reasoning segments, this field is optioanl
  • tool_calls: tool invocation plan
  • content: natural language output
These extend the standard OpenAI format.

​
4. Example Request (MiniMax-M2)

{
  "model": "minimax/minimax-m2",
  "messages": [
    {
      "role": "user",
      "content": "How's the weather in San Francisco?"
    },
    {
      "role": "assistant",
      "name": "MiniMax AI",
      "content": "",
      "tool_calls": [
        {
          "id": "call_function_asqvfevfc8af_1",
          "type": "function",
          "function": {
            "name": "get_weather",
            "arguments": "{\"location\": \"San Francisco, US\"}"
          }
        }
      ],
      "reasoning_details": [
        {
          "type": "reasoning.text",
          "id": "reasoning-text-1",
          "format": "openai-responses-v1",
          "index": 0,
          "text": "The user is asking about the weather in San Francisco..."
        }
      ]
    },
    {
      "role": "tool",
      "tool_call_id": "call_function_asqvfevfc8af_1",
      "content": "24℃, sunny"
    }
  ],
  "stream": true,
  "reasoning_split": true,
  "max_tokens": 1024,
  "temperature": 0.7,
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get weather for a specific location",
        "parameters": {
          "type": "object",
          "properties": {
            "location": { "type": "string" }
          },
          "required": ["location"]
        }
      }
    }
  ]
}

​
5. Example Response (Non-Streaming)

{
  "id": "07a4dedfdb1498b045498dfd42497639",
  "object": "chat.completion",
  "created": 1764303147,
  "model": "MiniMax",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "",
        "name": "MiniMax",
        "tool_calls": [
          {
            "index": 0,
            "id": "call_function_9w7wq1j9zmpl_1",
            "type": "function",
            "function": {
              "name": "get_weather",
              "arguments": "{\"location\": \"ShangHai\"}"
            }
          }
        ],
        "reasoning_content": "The user asked for Shanghai weather...",
        "reasoning_details": [
          {
            "type": "reasoning.text",
            "text": "The user is asking about the weather in Shanghai...",
            "id": "reasoning-text-1",
            "format": "openai-responses-v1",
            "index": 0
          }
        ]
      },
      "finish_reason": "tool_calls"
    }
  ]
}

​
6. Streaming Response Example

{
  "id": "664c5ad870c1888fbcbd267d9829e354",
  "object": "chat.completion.chunk",
  "created": 1764303504,
  "model": "minimax-m2",
  "choices": [
    {
      "index": 0,
      "delta": {
        "role": "assistant",
        "reasoning_content": "...\n\nThe user has specifically asked...",
        "reasoning_details": [
          {
            "type": "reasoning.text",
            "text": ".\n\nThe user has specifically asked...",
            "id": "reasoning-text-1",
            "format": "openai-responses-v1",
            "index": 0
          }
        ]
      },
      "finish_reason": null
    }
  ]
}

​
7. Developer Notes

​
7.1 Models That Support Interleaved Thinking

All models exposing reasoning_details through OpenAI-compatible APIs, including:
  • MiniMax-M2
  • (Upcoming) Novita Reasoning Series
  • Other reasoning-enabled partner models

​
7.2 Pricing

Billing is based on reasoning tokens, following the model’s pricing rules. reasoning_details will increase token usage.

​
7.3 Error Handling

You may encounter:
  • Missing tool parameters
  • Recursive or repeated tool calls
  • Incorrect assumptions in the reasoning phase
Ensure your application validates tool arguments and handles model errors gracefully.

​
8. Best Practices

✓ Always include full model messages in the next request Include:
  • content
  • tool_calls
  • reasoning_details
✓ Enable streaming for long-chain reasoning Streaming allows the client to:
  • Monitor the reasoning process
  • Detect incorrect tool plans early
  • Provide faster user feedback
✓ Combine with server-side state machines for stability For production systems, we recommend pairing Interleaved Thinking with deterministic guardrails, e.g.:
  • Parameter validators
  • Execution sandbox
  • Maximum recursion safeguards

​
9. Summary

Interleaved Thinking significantly enhances multi-step reasoning and tool-use reliability:
  • Transparent and inspectable step-wise reasoning
  • Adaptive planning between tool invocations
  • Stronger context retention across long workflows
  • Fully compatible with OpenAI-style Chat Completions API
Novita will continue expanding support for advanced reasoning models, bringing agent-level intelligence to the API layer.