How do I get JSON output from an Ollama model?

Add "format": "json" to the request. The model then returns a valid JSON object. For a specific shape, pass a JSON Schema in format instead of the plain string.

What is the difference between JSON mode and a JSON schema?

format: "json" only guarantees the output parses as JSON. A JSON Schema in format constrains the exact keys, types, and required fields, so the result matches a structure your code expects.

Does Ollama support GBNF grammar files?

Ollama's public API constrains output with a JSON Schema passed in format , not GBNF grammar files. For structured data, define a schema rather than reaching for a raw grammar.

Why does my model still return invalid or wrong-shaped JSON?

Loose format: "json" does not enforce keys. Pass a JSON Schema, set temperature to 0, and name the required fields in the prompt so a small model cannot drift from the shape.

Build / AI / Ollama

Structured JSON Outputs

Force reliable JSON from local models using the format flag and a JSON Schema for exact shapes.

TL;DR

Pass format: json to force a valid JSON object.
Constrain shape by passing a full JSON Schema in format.
Describe the exact keys you expect in the SYSTEM prompt.

Jump to section

JSON Mode
Schema-Constrained Output
Python Example
Determinism
Tips
Warnings
In Practice
FAQ

JSON Mode

format json

Force the response to be a valid JSON object.

curl http://localhost:11434/api/generate -d '{
  "model": "llama3.1",
  "prompt": "List 2 colors as JSON",
  "format": "json",
  "stream": false
}'

Prompt For Keys

Name the exact keys you want in the prompt.

"prompt": "Return name and age as JSON"

Parse Safely

The whole response body parses as JSON.

# response text is valid JSON to parse

Schema-Constrained Output

format Schema

Pass a JSON Schema object in the format field.

"format": { "type": "object", ... }

type object

Define the top-level type and properties.

{ "type": "object",
  "properties": { } }

required

List required keys the model must include.

"required": ["name", "age"]

Python Example

Pass format

Send a schema in the chat request's format.

ollama.chat(model="llama3.1",
  messages=msgs, format=schema)

Pydantic Schema

Generate a schema from a Pydantic model.

schema = Person.model_json_schema()

Validate

Parse the reply back into your model.

Person.model_validate_json(reply)

Determinism

temperature 0

Zero temperature removes sampling randomness.

# options: { "temperature": 0 }

Fixed seed

A set seed repeats output for a prompt.

# options: { "seed": 42 }

Schema + temp 0

Combine both for repeatable structured output.

# fixed shape + deterministic values

Tips

Pass a JSON Schema in format to pin exact keys and types, which is stricter and safer than plain format: "json" alone.
Also name the required fields in the prompt, since telling the model what to produce plus a schema yields the most reliable output.

Warnings

Plain format: "json" guarantees valid JSON but not your shape; pass a JSON Schema when specific keys are required.
Ollama constrains output with a JSON Schema, not GBNF grammar files; put the schema in the format field, not a grammar.

In Practice

Send a JSON Schema in the format field so the model returns exactly the keys your pipeline expects.

The format field carries a JSON Schema, not just the string json.
properties fixes the exact keys and their types.
required forces the model to include every listed key.
stream: false returns one complete JSON object to parse.

curl http://localhost:11434/api/chat -d '{
  "model": "llama3.1",
  "messages": [
    {"role":"user","content":"Alan Turing, 1912"}
  ],
  "stream": false,
  "format": {
    "type": "object",
    "properties": {
      "name": {"type": "string"},
      "born": {"type": "integer"}
    },
    "required": ["name", "born"]
  }
}'