Classify API

POST https://api.pinference.ai/api/v1/classify runs Qwen3 Reranker 0.6B on prompts you format yourself. For each prompt it returns the model’s probability of answering “yes” over “no”, and the logit behind it. The Rerank API builds the same prompts from one query and a list of documents. Use classify to score pairs with different queries in one request, to write your own question, or to work with logits.

Prompt format

Qwen publishes this prompt for the model. {instruction}, {query} and {document} are your text:

<|im_start|>system
Judge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be "yes" or "no".<|im_end|>
<|im_start|>user
<Instruct>: {instruction}
<Query>: {query}
<Document>: {document}<|im_end|>
<|im_start|>assistant
<think>

</think>

The prompt ends with two newlines after </think>. The Rerank API uses exactly this template, with the default instruction below. This function builds the same string:

DEFAULT_INSTRUCTION = "Given a web search query, retrieve relevant passages that answer the query"


def format_pair(query: str, document: str, instruction: str = DEFAULT_INSTRUCTION) -> str:
    return (
        "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query "
        'and the Instruct provided. Note that the answer can only be "yes" or "no".<|im_end|>\n'
        "<|im_start|>user\n"
        f"<Instruct>: {instruction}\n<Query>: {query}\n<Document>: {document}"
        "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
    )

The API sends each prompt to the model unchanged and reads the model’s preference for “yes” over “no” at the prompt’s last token. It doesn’t check the format. A prompt that asks a yes-or-no question and ends with the exact assistant suffix, <|im_start|>assistant\n<think>\n\n</think>\n\n, gets a meaningful score. Before the suffix you can change the system message, the field labels and the instruction. A prompt without the suffix still returns a score, but that score has no defined meaning.

Classify prompts

Send one prompt as a string, or up to 1,000 as an array. This example uses format_pair from above:

import os
import requests

prompts = [
    format_pair("Which planet is known as the Red Planet?", "Mars is often referred to as the Red Planet."),
    format_pair("Which planet is known as the Red Planet?", "Venus is often called Earth's twin because of its similar size."),
]
response = requests.post(
    "https://api.pinference.ai/api/v1/classify",
    headers={"Authorization": f"Bearer {os.environ['PRIME_API_KEY']}"},
    json={"model": "qwen/qwen3-reranker-0.6b", "input": prompts},
    timeout=120,
)
response.raise_for_status()
for item in response.json()["data"]:
    print(item["index"], item["label"], round(item["probs"][0], 3))

The request returns:

{
  "id": "req_...",
  "object": "list",
  "created": 1794009600,
  "model": "qwen/qwen3-reranker-0.6b",
  "data": [
    {"index": 0, "label": "yes", "probs": [0.997], "num_classes": 1, "logit": 5.73},
    {"index": 1, "label": "no", "probs": [0.025], "num_classes": 1, "logit": -3.65}
  ],
  "usage": {"prompt_tokens": 183, "total_tokens": 183, "completion_tokens": 0}
}

With curl, the prompt is a JSON string with its newlines and quotes escaped:

export PRIME_API_KEY="your-api-key"

curl https://api.pinference.ai/api/v1/classify \
  -H "Authorization: Bearer $PRIME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3-reranker-0.6b",
    "input": "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n<|im_start|>user\n<Instruct>: Given a web search query, retrieve relevant passages that answer the query\n<Query>: Which planet is known as the Red Planet?\n<Document>: Mars is often referred to as the Red Planet.<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
  }'

To charge a team, add -H "X-Prime-Team-ID: your-team-id" to each request. See team billing.

Request body

Field	Type	Description
`model`	string	`qwen/qwen3-reranker-0.6b`.
`input`	string or array of strings	1 to 1,000 prompts. Each must be a non-empty string of at most 32,768 tokens. A longer prompt is refused with a 400, and nothing is truncated.

The whole request body can be at most 16 MiB.

Response

Field	Description
`id`	The request’s ID, the same as the `X-Inference-Id` response header. Share it with support to trace a request.
`object`	Always `list`.
`created`	When the response was made, in Unix seconds.
`model`	The model ID from the request.
`data`	One entry per prompt, in request order.
`data[].index`	The prompt’s position in `input`, counting from 0.
`data[].probs`	A one-element array with the probability, from 0 to 1, of “yes” over “no”.
`data[].label`	`yes` when that probability is at least 0.5, otherwise `no`.
`data[].logit`	The model’s raw score. The probability is its sigmoid.
`data[].num_classes`	Always 1. The model has one output.
`usage.prompt_tokens`	The tokens the model read, summed over all prompts.
`usage.total_tokens`	The same as `prompt_tokens`.
`usage.completion_tokens`	Always 0. Nothing is generated.

Pick your own threshold for decisions. At a probability of 0.5 the model is undecided between “yes” and “no”. When two probabilities both round to 1.0, compare their logits to order them.

Usage and billing

Classification is billed on input tokens, usage.prompt_tokens, at $0.20 per million. There is no output charge.

Errors

Most errors return an error object with message, type and code, as the Chat Completions API does. A 401, a 404, or a 503 while authentication is unavailable returns a detail message instead.

Status	Cause
400	`Invalid classify request (input).` means `input` is missing or empty, holds an empty string or a non-string, or holds a prompt longer than 32,768 tokens. `Invalid classify request (too_many_inputs).` means more than 1,000 prompts.
401	The API key is missing or invalid.
402	The billed account has insufficient funds. See troubleshooting.
404	The request didn’t name a reranker, or its body isn’t JSON. A misspelled model ID, a missing `model` and a non-JSON body all return a `detail` of `Not Found`. This check comes before the API key check, so a wrong model with a bad key returns 404.
413	The request body is larger than 16 MiB.
429	The model has no free capacity. Retry after the number of seconds in the `Retry-After` header.
502 or 503	The model failed to answer. Retry the request.

vLLM compatibility

The request and response follow vLLM’s classification API, with logit added.

API Documentation

Compute API

Inference API

Traces API

Availability

Billing

Disks

evals

factory

Feedback

hosted-evaluations

Images

inference

metrics

Pods

Secrets

slurm-clusters

SSH Keys

teams

API Keys

training

Tunnel

user

Prompt format

Classify prompts

Request body

Response

Usage and billing

Errors

vLLM compatibility

​Prompt format

​Classify prompts

​Request body

​Response

​Usage and billing

​Errors

​vLLM compatibility

Prompt format

Classify prompts

Request body

Response

Usage and billing

Errors

vLLM compatibility