Model compression

Compress models using techniques like quantization to reduce inference costs.

Overview

The Model Compression API compresses custom full-parameter fine-tuned models through techniques such as quantization, reducing inference memory footprint and improving throughput. Currently, the compression feature only supports quantization and covers the complete lifecycle from querying templates, creating tasks, polling status, retrieving logs, to canceling/deleting tasks.

Typical flow:

• List quantizable models and configuration templates → obtain the template_id and the quantizable model
• Create a compression task → obtain the job_id
• Poll Query compression task / Get compression task logs → until SUCCEEDED / FAILED / CANCELED
• After SUCCEEDED, use quantized_output to create a deployment; when no longer needed, Cancel compression task / Delete compression task

Domain for all API endpoints: https://dashscope.aliyuncs.com. Authentication uniformly uses Authorization: Bearer ${YOUR_API_KEY}, and POST requests must include Content-Type: application/json.

For the meanings of task object fields and the state machine, see Compression task object; for unified error codes, see Error codes at the end of the document.

Quick Start

The model compression API is currently available only in the Beijing Region. If you use another Region, complete model compression operations through the Bailian console of that Region.

The model compression API provides a complete set of RESTful interfaces covering query templates, creating jobs, polling status, obtaining logs, and canceling/deleting jobs. This document is intended for developers to integrate compression capabilities via OpenAPI or SDK. For console introductions, see related documents.

Prerequisites

Before calling the interfaces in this document, please complete the following:

1. Activated Alibaba Cloud Bailian service and completed real-name verification.
2. The current workspace has at least one custom full-parameter fine-tuned model based on qwen3.5-flash-2026-02-23 (completed via the fine-tuning job interface). The current compression feature only supports this model; LoRA models and already-quantized models are not supported.
3. Obtained an API Key (see Obtain API Key).

The deployment unit specifications supported by the compressed output model are determined by the selected quantization template, and the deployment quantity is configured under Dedicated Deployment in the Bailian console. The current compression feature is free for a limited time.

Interface List

All interface domains: https://dashscope.aliyuncs.com

Authentication

All interfaces carry the API Key via HTTP Header:

Authorization: Bearer ${YOUR_API_KEY}

Content-Type: application/json applies to POST request body scenarios.

Get Started in 5 Minutes

pip install requests

import requests, time

API_KEY = "YOUR_API_KEY"
BASE = "https://dashscope.aliyuncs.com/api/v1/fine-tunes/compress"
HEADERS = {"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json"}

# 1. Create compression job
resp = requests.post(f"{BASE}/jobs", headers=HEADERS, json={
    "model": "qwen3.5-flash-2026-02-23-ft-***",  # Custom fine-tuned model ID
    "template_id": "quant-flash-nvfp4-mlp-nomtp",          # Obtain via GET /templates
    "output_model_suffix": "test",                          # Max 8 characters, lowercase letters and digits only
}).json()
job_id = resp["output"]["job_id"]
print("Job:", job_id)

# 2. Poll until terminal state
status = resp["output"]["status"]
while status not in ("SUCCEEDED", "FAILED", "CANCELED"):
    time.sleep(30)
    resp = requests.get(f"{BASE}/jobs/{job_id}", headers=HEADERS).json()
    status = resp["output"]["status"]
    print("Status:", status)

# 3. Process results
if status == "SUCCEEDED":
    print("Quantized model:", resp["output"]["quantized_output"])
    # Use quantized_output to call the model deployment interface for deployment
elif status == "FAILED":
    print("Error:", resp["output"]["error"])

# List all successfully compressed jobs
resp = requests.get(f"{BASE}/jobs", headers=HEADERS, params={"status": "SUCCEEDED", "page_size": 20}).json()
for j in resp["output"]["jobs"]:
    print(j["job_id"], j["template_name"], j["quantized_output"])

# Get job logs
resp = requests.get(f"{BASE}/jobs/{job_id}/logs", headers=HEADERS, params={"offset": 0, "line": 100}).json()
for line in resp["output"]["logs"]:
    print(line)

Output model naming rules:

quantized_output = {base_model}-{output_model_suffix}-{job_id}

For example, base_model is qwen3.5-flash-2026-02-23, suffix is test, job_id is quant-202604111200-a1b2, and the output model ID is:

qwen3.5-flash-2026-02-23-test-quant-202604111200-a1b2

See each interface detail page for cURL usage of each interface.

Compression job object

The model compression API is currently available only in the Beijing Region. If you use another Region, complete model compression in that Region's Bailian console.

Object properties

Job status

Status transition:

PENDING ─→ QUEUING ─→ RUNNING ─→ SUCCEEDED
                  │           │
                  ↓           ↓
                CANCELING ─→ FAILED / CANCELED

List quantizable models and configuration templates

Lists all quantizable custom fine-tuned models of the current user, as well as the compression templates bound to each model. Templates are bound to models; different combinations of model architecture × precision × target MU specifications correspond to different templates.

Only returns custom models that the current user has fully fine-tuned (SFT/DPO/CPT) based on a base model. LoRA fine-tuned models and already-quantized models will not appear in the results.

GET /api/v1/fine-tunes/compress/templates

curl "https://dashscope.aliyuncs.com/api/v1/fine-tunes/compress/templates" \
  -H "Authorization: Bearer ${API_KEY}"

{
  "request_id": "uuid-string",
  "output": {
    "base_models": ["qwen3.5-flash-2026-02-23"],
    "custom_models": [
      {
        "model": "qwen3.5-flash-2026-02-23-ft-***",
        "model_name": "My SFT fine-tuned model",
        "base_model": "qwen3.5-flash-2026-02-23",
        "templates": [
          {
            "template_id": "quant-flash-nvfp4-mlp-nomtp",
            "template_name": "W4A4 NVFP4 high-performance compression-MU5/MU8/MU9",
            "description": "Balances high precision and high performance under lower-bit compression, further reducing memory usage and improving inference throughput.",
            "compress_type": "quantization",
            "hyper_parameters": []
          }
        ]
      }
    ]
  }
}

{
  "request_id": "uuid-string",
  "output": {
    "base_models": ["qwen3.5-flash-2026-02-23"],
    "custom_models": [
      {
        "model": "qwen3.5-flash-2026-02-23-ft-***",
        "model_name": "My SFT fine-tuned model",
        "base_model": "qwen3.5-flash-2026-02-23",
        "templates": [
          {
            "template_id": "quant-flash-nvfp4-mlp-nomtp",
            "template_name": "W4A4 NVFP4 high-performance compression-MU5/MU8/MU9",
            "description": "Balances high precision and high performance under lower-bit compression, further reducing memory usage and improving inference throughput.",
            "compress_type": "quantization",
            "hyper_parameters": [
              {
                "name": "calib_input",
                "type": "string",
                "display_name": "Calibration input",
                "description": "Whether to enable calibration input",
                "support_values": ["true"],
                "defaultValue": "true",
                "recommend_value": "true",
                "required": false
              }
            ]
          }
        ]
      }
    ]
  }
}

Create a compression job

POST /api/v1/fine-tunes/compress/jobs

curl -X POST "https://dashscope.aliyuncs.com/api/v1/fine-tunes/compress/jobs" \
  -H "Authorization: Bearer ${API_KEY}" \
  -H "Content-Type: application/json" \
  -d '{
    "job_name": "qwen3.5-flash compression job",
    "model": "qwen3.5-flash-2026-02-23-ft-***",
    "template_id": "quant-flash-nvfp4-mlp-nomtp",
    "custom_calibration_file_ids": ["file-***"],
    "output_model_suffix": "test"
  }'

{
  "request_id": "uuid-string",
  "output": {
    "job_id": "quant-202604111200-a1b2",
    "job_name": "qwen3.5-flash compression job",
    "status": "PENDING",
    "model": "qwen3.5-flash-2026-02-23-ft-***",
    "base_model": "qwen3.5-flash-2026-02-23",
    "template_id": "quant-flash-nvfp4-mlp-nomtp",
    "template_name": "W4A4 NVFP4 high-performance compression-MU5/MU8/MU9",
    "training_type": "quantization",
    "compress_type": "quantization",
    "hyper_parameters": {},
    "custom_calibration_file_ids": ["file-***"],
    "quantized_output": null,
    "create_time": "2026-04-11 12:00:00",
    "start_time": null,
    "end_time": null,
    "error": null,
    "group": "quantization"
  }
}

Response parameters: Field meanings are the same as the request parameters.

List compression jobs

Supports filtering by status, model, template, quantization spec, algorithm, time range, job name/ID, and supports sorting by creation time and pagination.

GET /api/v1/fine-tunes/compress/jobs

# Combined search: by status + algorithm + pagination
curl "https://dashscope.aliyuncs.com/api/v1/fine-tunes/compress/jobs?status=SUCCEEDED&quant_method=gptq&page_size=10" \
  -H "Authorization: Bearer ${API_KEY}"

# Time range + search by job name + ascending by create time
curl "https://dashscope.aliyuncs.com/api/v1/fine-tunes/compress/jobs?start_time=2026-04-01&end_time=2026-04-30&search_key=qwen3&select_key=job_name&sort_by=create_time&sort_order=asc" \
  -H "Authorization: Bearer ${API_KEY}"

{
  "request_id": "uuid-string",
  "output": {
    "total": 42,
    "page_no": 1,
    "page_size": 10,
    "jobs": [
      {
        "job_id": "quant-202604111200-a1b2",
        "job_name": "qwen3.5-flash compression job",
        "status": "SUCCEEDED",
        "model": "qwen3.5-flash-2026-02-23-ft-***",
        "base_model": "qwen3.5-flash-2026-02-23",
        "template_id": "quant-flash-nvfp4-mlp-nomtp",
        "template_name": "W4A4 NVFP4 high-performance compression-MU5/MU8/MU9",
        "training_type": "quantization",
        "compress_type": "quantization",
        "custom_calibration_file_ids": ["file-***"],
        "quantized_output": "qwen3.5-flash-2026-02-23-test-quant-202604111200-a1b2",
        "create_time": "2026-04-11 12:00:00",
        "start_time": "2026-04-11 12:02:30",
        "end_time": "2026-04-11 13:02:30",
        "group": "quantization",
        "usage": 3600
      }
    ]
  }
}

Query a compression job

GET /api/v1/fine-tunes/compress/jobs/{job_id}

curl "https://dashscope.aliyuncs.com/api/v1/fine-tunes/compress/jobs/quant-202604111200-a1b2" \
  -H "Authorization: Bearer ${API_KEY}"

{
  "request_id": "uuid-string",
  "output": {
    "job_id": "quant-202604111200-a1b2",
    "job_name": "qwen3.5-flash compression job",
    "job_description": "...",
    "status": "SUCCEEDED",
    "model": "qwen3.5-flash-2026-02-23-ft-***",
    "base_model": "qwen3.5-flash-2026-02-23",
    "template_id": "quant-flash-nvfp4-mlp-nomtp",
    "template_name": "W4A4 NVFP4 high-performance compression-MU5/MU8/MU9",
    "template_description": "Maintains high accuracy and high performance under lower-bit compression, further reducing GPU memory usage and improving inference throughput.",
    "training_type": "quantization",
    "compress_type": "quantization",
    "hyper_parameters": {},
    "custom_calibration_file_ids": ["file-***"],
    "quantized_output": "qwen3.5-flash-2026-02-23-test-quant-202604111200-a1b2",
    "create_time": "2026-04-11 12:00:00",
    "start_time": "2026-04-11 12:02:30",
    "end_time": "2026-04-11 13:02:30",
    "error": null,
    "group": "quantization",
    "usage": 3600
  }
}

Get compression job logs

GET /api/v1/fine-tunes/compress/jobs/{job_id}/logs

Where {job_id} is the compression job ID, which can be obtained via the create compression job or list compression jobs interface.

curl "https://dashscope.aliyuncs.com/api/v1/fine-tunes/compress/jobs/quant-202604111200-a1b2/logs?offset=0&line=50" \
	-H "Authorization: Bearer ${API_KEY}"

{
  "request_id": "uuid-string",
  "output": {
    "total": 15,
    "logs": [
      "2026-04-11 12:02:35 - INFO - Starting quantization...",
      "2026-04-11 12:30:00 - INFO - Quantization progress: 100%",
      "2026-04-11 12:35:00 - INFO - Quantization succeeded!"
    ]
  }
}

• When a custom calibration dataset is provided for the job, the logs contain a data processing completion marker data process succeeded, start to quantization
• The log interface has filtered out internal system markers and only returns user-readable compression progress information

Cancel a compression job

You can cancel only jobs in the PENDING, QUEUING, or RUNNING state. Cancellation is an asynchronous operation. The job first enters the CANCELING transitional state and eventually becomes CANCELED.

POST /api/v1/fine-tunes/compress/jobs/{job_id}/cancel

Where {job_id} is the compression job ID, which can be obtained from the Create compression job or List compression jobs API.

curl -X POST "https://dashscope.aliyuncs.com/api/v1/fine-tunes/compress/jobs/quant-202604111200-a1b2/cancel" \
  -H "Authorization: Bearer ${API_KEY}"

{
  "request_id": "uuid-string",
  "output": { "status": "success" }
}

Delete a compression job

You can delete only jobs in a terminal state (SUCCEEDED / FAILED / CANCELED). Deleting a job record does not delete the already-generated quantized model (quantized_output).

DELETE /api/v1/fine-tunes/compress/jobs/{job_id}

Where {job_id} is the compression job ID, which can be obtained from the Create compression job or List compression jobs API.

curl -X DELETE "https://dashscope.aliyuncs.com/api/v1/fine-tunes/compress/jobs/quant-202604111200-a1b2" \
  -H "Authorization: Bearer ${API_KEY}"

{
  "request_id": "uuid-string",
  "output": { "status": "success" }
}

Error code

Common error codes

Business error codes

The following business error codes are listed by scenario. External Code is the actual code field value returned by the interface.

Error response example

{
  "request_id": "uuid-string",
  "code": "InvalidParameter",
  "message": "The specified model 'xxx-lora-yyy' is a LoRA model and not supported for quantization."
}