Inworld
Inworld connects your flows to Inworld Inference, Inworld’s model service. With one credential you get the models Inworld hosts, built for fast conversation (Gemma 4, GLM, DeepSeek, Qwen, MiMo), and over a hundred models from other providers such as OpenAI, Anthropic, Google, Groq or xAI. You can use it in two ways: as an agent’s model, connecting its Model output to the agent’s Model input, or on its own, answering the text that reaches it through Input.
When to use it
- Give an agent, or a voice agent, the model it understands requests with, decides which tools to use and answers with.
- Try models from different providers without a credential for each one, on a single bill.
- Write, summarize or classify texts inside a flow, with no agent needed.
- Process a list of texts one by one, for example every review of a month, with Batch Mode.
Configuration
In the Common tab:
| Field | What it does |
|---|---|
| Model Name (required) | The model that answers. Once you pick the credential, the models your key can use are loaded: Inworld’s own first (inworld/models/…), then other providers’ (openai/…, anthropic/…). Click the refresh button next to the field to load them again. |
| Temperature | How varied the answers are: low values make them steadier and more predictable; high values, more creative. Default: 0.1. |
| Credential (required) | The Inworld API credential, the same one the Inworld voice and transcription nodes use. Pick one from the list (Select a credential) or create another with Create new credential and paste the key exactly as you copy it from Inworld. See Credentials. |
| Connection | What to connect |
|---|---|
| Input | The text the model answers when it works on its own, for example a Chat Input message or a Prompt’s result. |
Advanced options
In the Advanced tab:
| Field | What it does |
|---|---|
| System Message | General instructions for the model when it works on its own: its role, the tone, what to avoid. You can type them or connect them from another component. |
| Stream | Shows the answer in the chat as it is written instead of waiting for it to finish. Only affects the chat. Default: off. |
| Reasoning Effort | How much the model thinks before answering. The list shows the levels the selected model accepts. default leaves it to the model, except in a voice agent, where it means the fastest level the model allows. Higher levels take longer. Default: default. |
| Max Tokens | The maximum length of the answer, in tokens (pieces of words). 0 or empty sets no limit of its own. |
| Seed | A fixed number that helps get similar answers when you repeat the same request. Default: 1. |
| Other options | Model Kwargs: fine-tuning for advanced users. Leave it as it is if you don’t need it. |
Batch processing
With the Batch Mode switch on, the component runs once for each item of the list you connect to Batch Input and returns all results together in a list. Parallel Executions sets how many items run at once (1 to 5). Skip if Empty Batch Input skips the component when the list arrives empty, and Skip Branch if Empty Output stops the following components when the result is empty. More in Key concepts.
Result
| Output | What it holds |
|---|---|
| Model | The configured model. It appears at the top of the component. Connect it to the Model input of an agent, a Voice Agent or any other component that asks for a model. |
| Text | The model’s answer to the text received in Input. Connect it, for example, to a Chat Output or another component that works with text. |
The two ways of using it exclude each other: when you connect the Model output, the component hides Input and Text, because the model now works inside the other component; and if you connect Input, Model is hidden.
Example
An online shop answers the phone with a voice agent that speaks with an Inworld voice. It connects Inworld to the agent’s Model input with the same credential, picks inworld/models/gemma-4-31b-it and leaves Reasoning Effort at default so it answers without pauses.
Best practices
- Use a low Temperature when you need exact answers, such as classifying or extracting data, and raise it for more creative texts.
- When the model feeds an agent, configure only the model here: instructions and tools are set on the agent.
- Inworld serves its text models from the United States: from Europe each answer adds about a tenth of a second of network (measured from a server in Germany on 15 October 2026).
Security
- An agent’s tools need a payment method on your Inworld workspace; with no credit, Inworld refuses the requests.
- Keep the key in the credential only: never type it into a field of the flow.