MaxCompute:MaxFrame AI Function

Model catalog (continuously updated)

Qwen 3 series: Inference-optimized. Supports multilingual translation, complex text generation, and code generation. Best for scenarios that require high-precision output.

Qwen Embedding models: Designed for vectorization tasks. Converts text to vectors efficiently. Best for semantic search and similarity matching.

DeepSeek-R1-Distill-Qwen series: Lightweight, compressed using knowledge distillation. Best for fast inference in resource-constrained environments.

generate — general-purpose interface

Use generate when you need custom prompt templates and parameter control.

from maxframe.learn.contrib.llm.models.managed import ManagedTextLLM

llm = ManagedTextLLM(name="<model_name>")

messages = [
    {"role": "system", "content": "system_messages"},
    {"role": "user", "content": "user_messages"},
]

result_df = llm.generate(<df>, prompt_template=messages)
print(result_df.execute())

Parameters:

Task-specific interfaces — GU resources only

Use task-specific interfaces for common scenarios without writing a prompt. No prompt engineering required — pass your data and let the interface handle the rest. Currently supported: translate and extract.

Task-specific interfaces require GU (GPU Unit) compute resources.

`translate` example:

from maxframe.learn.contrib.llm.models.managed import ManagedTextLLM

llm = ManagedTextLLM(name="<model_name>")

translated_df = llm.translate(
    df["english_column"],
    source_language="english",
    target_language="Chinese",
    examples=[("Hello", "你好"), ("Goodbye", "再见")],
)

translated_df.execute()

`extract` example:

from maxframe.learn.contrib.llm.models.managed import ManagedTextLLM
from pydantic import BaseModel
from typing import List, Optional

llm = ManagedTextLLM(name="<model_name>")

class Record(BaseModel):
    field_a: str
    field_b: Optional[str] = None

result_df = llm.extract(
    df["text_column"],
    description="Extract structured data from the record. Return strict JSON matching the schema.",
    schema=Record
)
result_df.execute()

Translate contracts at scale

Scenario: A multinational corporation needs to translate 100,000 English contracts into Chinese and annotate key clauses. This example covers:

1. Create a MaxCompute source table with contract data

2. Define a translation prompt with a few-shot example

3. Run batch inference and write results back to MaxCompute

Step 1: Create the source table

CREATE TABLE IF NOT EXISTS raw_contracts (
    id BIGINT,
    en STRING
);

-- Insert sample data
INSERT INTO raw_contracts VALUES
(1, 'This Agreement is made and entered into as of the Effective Date by and between Party A and Party B.'),
(2, 'The Contractor shall perform the Services in accordance with the terms and conditions set forth herein.'),
(3, 'All payments shall be made in US Dollars within thirty (30) days of receipt of invoice.'),
(4, 'Either party may terminate this Agreement upon thirty (30) days written notice to the other party.'),
(5, 'Confidential Information shall not be disclosed to any third party without prior written consent.');

Step 2: Run inference

from maxframe.learn.contrib.llm.models.managed import ManagedTextLLM

# 1. Initialize the Qwen3-1.7B model
llm = ManagedTextLLM(name="Qwen3-1.7B")

# 2. Load the source data
df = md.read_odps_table("raw_contracts")

# 3. Define the translation prompt with a few-shot example
messages = [
    {
        "role": "system",
        "content": "You are a document translation expert who can fluently translate English text into Chinese.",
    },
    {
        "role": "user",
        "content": "Translate the following English text into Chinese. Output only the translated text.\n\nExample:\nInput: Hi\nOutput: Hello.\n\nText to translate:\n\n{en}",
    },
]

# 4. Run translation using the generate interface, referencing the `en` column
result_df = llm.generate(
    df,
    prompt_template=messages,
    params={
        "temperature": 0.7,
        "top_p": 0.8,
    },
).execute()

# 5. Write results back to MaxCompute
result_df.to_odps_table("raw_contracts_result").execute()

Extract structured data from medical records

Scenario: Medical consultation records contain unstructured text. This example covers:

1. Load records from a MaxCompute table

2. Define a structured schema using Pydantic

3. Run extraction using the extract interface

4. Write structured results back to MaxCompute

Step 1: Create the source table

CREATE TABLE IF NOT EXISTS traditional_chinese_medicine (
    index BIGINT,
    text STRING
);

-- Insert sample data
INSERT INTO traditional_chinese_medicine VALUES
(1, 'Patient Zhang, male, 45 years old. Chief complaint: recurrent cough for 2 weeks. Current symptoms: cough with abundant white and sticky phlegm, chest tightness, shortness of breath, poor appetite, and loose stools. Tongue: white and greasy coating. Pulse: slippery. Diagnosis: Phlegm-dampness obstructing the lungs. Treatment principle: Dry dampness, resolve phlegm, regulate qi, and stop coughing. Prescription: Modified Erchen Tang.'),
(2, 'Patient Li, female, 32 years old. Visited for "insomnia and excessive dreaming for 1 month." Accompanied by heart palpitations, forgetfulness, mental fatigue, and a sallow complexion. Tongue: pale with a thin white coating. Pulse: thin and weak. Diagnosis: Heart and spleen deficiency syndrome. Prescription: Modified Guipi Tang, one dose daily, decocted in water.'),
(3, 'Patient Wang, 68 years old. Chief complaint: lower back and knee soreness, frequent nocturia for half a year. Accompanied by aversion to cold, cold limbs, tinnitus like cicadas, and listlessness. Tongue: pale, swollen with teeth marks, and a white, slippery coating. Pulse: deep and thin. TCM diagnosis: Kidney yang deficiency syndrome. Treatment: Warm and supplement kidney yang. Prescription: Modified Jinkui Shenqi Wan.'),
(4, 'Patient Zhao, 5 years old. Fever for 3 days, highest temperature 39.5°C, slight aversion to wind and cold, nasal congestion with turbid discharge, red and swollen painful throat. Tongue: red tip with a thin yellow coating. Pulse: floating and rapid. Diagnosis: Wind-heat invading the lungs syndrome. Treatment principle: Dispel wind, clear heat, diffuse the lungs, and stop coughing. Prescription: Modified Yinqiao San.'),
(5, 'Patient Liu, male, 50 years old. Recurrent epigastric bloating and pain for 3 years, worsened for 1 week. Symptoms: epigastric fullness, frequent belching, symptoms worsen with emotional fluctuations, and irregular bowel movements. Tongue: red with a thin yellow coating. Pulse: wiry. Diagnosis: Liver-stomach disharmony syndrome. Treatment principle: Soothe the liver, harmonize the stomach, regulate qi, and relieve pain. Prescription: Modified Chaihu Shugan San combined with Zuojin Wan.');

Step 2: Run extraction

from maxframe.learn.contrib.llm.models.managed import ManagedTextLLM
from pydantic import BaseModel
from typing import List, Optional

# 1. Initialize the model (GU resource required for task interfaces)
llm = ManagedTextLLM(name="Qwen3-4B-Instruct-2507-FP8")

# 2. Load the data and distribute across four partitions for parallel processing
df = md.read_odps_table("traditional_chinese_medicine", index_col="index")
df = df.mf.rebalance(num_partitions=4)

# 3. Define the target schema
class MedicalRecord(BaseModel):
    """Structured schema for a traditional Chinese medicine consultation record."""
    patient_name: Optional[str] = None     # Patient's name (e.g., "Zhang")
    age: Optional[int] = None              # Age
    gender: Optional[str] = None           # Gender ("Male"/"Female")
    chief_complaint: str                   # Chief complaint
    symptoms: List[str]                    # List of symptoms
    tongue: str                            # Tongue diagnosis
    pulse: str                             # Pulse diagnosis
    diagnosis: str                         # TCM syndrome diagnosis
    treatment_principle: str               # Treatment principle
    prescription: str                      # Prescription name

# 4. Extract structured data using the preset extract interface
result_df = llm.extract(
    df["text"],
    description="Extract structured data from the consultation record. Return strict JSON matching the schema.",
    schema=MedicalRecord
)
result_df.execute()

# 5. Write results to MaxCompute
result_df.to_odps_table("result").execute()

Choose between CU and GU

MaxFrame supports two compute resource types:

For models 8B and larger, GU resources deliver significantly better performance than CU. Task-specific interfaces (translate, extract) require GU resources.

Configure the resource type on your session:

# Use CU compute resources
options.session.quota_name = "mf_cpu_quota"

# Use GU compute resources
options.session.gu_quota_name = "mf_gpu_quota"

Scale with parallel inference

For large-scale inference jobs, use rebalance to pre-distribute data across worker nodes before running inference. Each worker independently loads the model, avoiding cold start delays from sequential loading.

# Distribute data across 4 partitions before inference
parallel_partitions = 4
df = df.mf.rebalance(num_partitions=parallel_partitions)

After inference completes, results are written to MaxCompute by partition and are available for downstream data analytics.

Sample data

Customer Credit Rating Investigation Report - Number: RJ22222222-ZZ
Customer Name: Mr. Wang
ID Card Number: 110000111111111111
Contact: 111000000000  
Permanent Address: Room 10000, Building 10000, Chaoyang Community, Chaoyang Street, Chaoyang District, Beijing  
**Asset Overview**  
- China Merchants Bank Account: 6200 0000 0000 0000 (Balance: ¥2,000,000.00)  
- Ping An Securities Account: A100000000 (Position Market Value: ¥1,000,000.00)  
**Recent Large Transactions**  
1. 2088-15-35 Transfer to Ms. Li (Account: 6200 0000 0000 0000) ¥1,500,000.00, Remarks: Down payment for house purchase - No. 10000, Lane 10000, Pudong New Road, Pudong New District, Shanghai.
2. 2088-15-35 Received payment of ¥4,000,000.00 from "Shenzhen XX Electronic Technology Co., Ltd." (Bank: Bank of China Shenzhen Branch).
**Approval Opinion**  
Customer credit rating AAA, recommended for maximum credit limit. Handled by: Ms. Zhang (Employee ID: ZX000000000), Contact: zhangkaixin@example.com / Extension: 88888.

Sample code

import os
from maxframe import new_session
from odps import ODPS
import pandas as pd
import maxframe.dataframe as md
from maxframe.learn.contrib.llm.models.managed import ManagedTextLLM

o = ODPS(
    os.getenv('ALIBABA_CLOUD_ACCESS_KEY_ID'),
    os.getenv('ALIBABA_CLOUD_ACCESS_KEY_SECRET'),
    project='MaxCompute_Project_Name',
    endpoint='https://service.cn-hangzhou.maxcompute.aliyun.com/api',
)

session = new_session(odps_entry=o)

# Do not truncate column text to display the full content.
pd.set_option("display.max_colwidth", None)
# Display all columns to prevent middle columns from being omitted.
pd.set_option("display.max_columns", None)

query_list = [
"Customer Credit Rating Investigation Report - Number: RJ22222222-ZZ Customer Name: Mr. Wang ID Card Number: 110000111111111111 Contact: 111000000000  Permanent Address: Room 10000, Building 10000, Chaoyang Community, Chaoyang Street, Chaoyang District, Beijing  **Asset Overview**  China Merchants Bank Account: 6200 0000 0000 0000 (Balance: ¥2,000,000.00)  Ping An Securities Account: A100000000 (Position Market Value: ¥1,000,000.00)  **Recent Large Transactions**  1. 2088-15-35 Transfer to Ms. Li (Account: 6200 0000 0000 0000) ¥1,500,000.00, Remarks: Down payment for house purchase - No. 10000, Lane 10000, Pudong New Road, Pudong New District, Shanghai. 2. 2088-15-35 Received payment of ¥4,000,000.00 from 'Shenzhen XX Electronic Technology Co., Ltd.' (Bank: Bank of China Shenzhen Branch). **Approval Opinion**  Customer credit rating AAA, recommended for maximum credit limit. Handled by: Ms. Zhang (Employee ID: ZX000000000), Contact: zhangkaixin@example.com / Extension 88888."
]

df = md.DataFrame({"query": query_list})

llm = ManagedTextLLM(name="DeepSeek-R1-0528-Qwen3-8B")

# DeepSeek models should not be provided with system prompts.
messages = [
    {
        "role": "user",
        "content": """Act as a text risk control assessor. Evaluate the safety of the following online text and provide reasons. You must meet the following requirements:

1. **Violation types to identify** (from highest to lowest priority):
   - ID card information
   - Account information
   - Phone number information
   - Transaction amount information
   - Name information
   - Address information
   - Number information

2. **Output format**:
    Use the following pure JSON format. Ensure the output does not contain ``` or any other Markdown formatting symbols.
    "reason": "Briefly explain the basis for the rating (within 100 words).",
    "category": "Main violation type (or 'Safe' if none)."

The following is the text to evaluate:
{query}
""",
    },
]

result_df = llm.generate(
    df,
    prompt_template=messages,
)

print(result_df.execute())

Sample data

Customer Credit Rating Investigation Report - Number: RJ22222222-ZZ
Customer Name: Mr. Wang
ID Card Number: 110000111111111111
Contact: 111000000000  
Permanent Address: Room 10000, Building 10000, Chaoyang Community, Chaoyang Street, Chaoyang District, Beijing  
**Asset Overview**  
- China Merchants Bank Account: 6200 0000 0000 0000 (Balance: ¥2,000,000.00)  
- Ping An Securities Account: A100000000 (Position Market Value: ¥1,000,000.00)  
**Recent Large Transactions**  
1. 2088-15-35 Transfer to Ms. Li (Account: 6200 0000 0000 0000) ¥1,500,000.00, Remarks: Down payment for house purchase - No. 10000, Lane 10000, Pudong New Road, Pudong New District, Shanghai.
2. 2088-15-35 Received payment of ¥4,000,000.00 from "Shenzhen XX Electronic Technology Co., Ltd." (Bank: Bank of China Shenzhen Branch).
**Approval Opinion**  
Customer credit rating AAA, recommended for maximum credit limit. Handled by: Ms. Zhang (Employee ID: ZX000000000), Contact: zhangkaixin@example.com / Extension: 88888.

Sample code

import os
from maxframe import new_session
from odps import ODPS
import pandas as pd
import maxframe.dataframe as md
from maxframe.learn.contrib.llm.models.managed import ManagedTextLLM

o = ODPS(
    os.getenv('ALIBABA_CLOUD_ACCESS_KEY_ID'),
    os.getenv('ALIBABA_CLOUD_ACCESS_KEY_SECRET'),
    project='MaxCompute_Project_Name',
    endpoint='https://service.cn-hangzhou.maxcompute.aliyun.com/api',
)

session = new_session(odps_entry=o)

# Do not truncate column text to display the full content.
pd.set_option("display.max_colwidth", None)
# Display all columns to prevent middle columns from being omitted.
pd.set_option("display.max_columns", None)

query_list = [
"Customer Credit Rating Investigation Report - Number: RJ22222222-ZZ Customer Name: Mr. Wang ID Card Number: 110000111111111111 Contact: 111000000000  Permanent Address: Room 10000, Building 10000, Chaoyang Community, Chaoyang Street, Chaoyang District, Beijing  **Asset Overview**  China Merchants Bank Account: 6200 0000 0000 0000 (Balance: ¥2,000,000.00)  Ping An Securities Account: A100000000 (Position Market Value: ¥1,000,000.00)  **Recent Large Transactions**  1. 2088-15-35 Transfer to Ms. Li (Account: 6200 0000 0000 0000) ¥1,500,000.00, Remarks: Down payment for house purchase - No. 10000, Lane 10000, Pudong New Road, Pudong New District, Shanghai. 2. 2088-15-35 Received payment of ¥4,000,000.00 from 'Shenzhen XX Electronic Technology Co., Ltd.' (Bank: Bank of China Shenzhen Branch). **Approval Opinion**  Customer credit rating AAA, recommended for maximum credit limit. Handled by: Ms. Zhang (Employee ID: ZX000000000), Contact: zhangkaixin@example.com / Extension 88888."
]

df = md.DataFrame({"query": query_list})

llm = ManagedTextLLM(name="DeepSeek-R1-0528-Qwen3-8B")

# DeepSeek models should not be provided with system prompts.
messages = [
    {
        "role": "user",
        "content": """Act as a text risk control assessor. Evaluate the safety of the following online text and provide reasons. You must meet the following requirements:

1. **Violation types to identify** (from highest to lowest priority):
   - ID card information
   - Account information
   - Phone number information
   - Transaction amount information
   - Name information
   - Address information
   - Number information

2. **Output format**:
    Use the following pure JSON format. Ensure the output does not contain ``` or any other Markdown formatting symbols.
    "reason": "Briefly explain the basis for the rating (within 100 words).",
    "category": "Main violation type (or 'Safe' if none)."

The following is the text to evaluate:
{query}
""",
    },
]

result_df = llm.generate(
    df,
    prompt_template=messages,
)

print(result_df.execute())