LevelField: Manufacturing Process Planning from Engineering Drawings

LevelField is an open-source, 9B-parameter multimodal model for non-standard machining process planning. It helps process engineers interpret 2D manufacturing drawings and draft machining routes with optimization suggestions.

Try the online demo · Download model files · Level AI

Model overview

Attribute	Details
Developer	Level AI (蓝沃AI)
Base model	Qwen3.5-9B
Parameters / precision	9B / BF16
Release format	Merged safetensors weights, approximately 20 GB, with inference configuration files
Languages	Chinese and English
Task	Drawing interpretation and manufacturing process planning
License	Apache 2.0; commercial use is permitted under the license terms

What the model does

Capability	Description
Drawing interpretation	Extracts part names, drawing numbers, materials, dimensions, tolerances, surface roughness, and technical requirements.
Process planning	Uses reviewed machining requirements to propose an operation sequence, identify unsuitable processes, and suggest improvements.
Process coverage	Supports the 18 manufacturing process types listed below.

The drawing workflow covers PDF and image documents. The API example below sends a JPEG image; render PDF pages to images before using this API workflow.

The process-planning prompt uses these 18 process types:

Process	Chinese label
Turning	车削
Milling	铣削
Drilling	钻削
Grinding	磨削
Casting	铸造
Forging	锻造
Heat treatment	热处理
Surface treatment	表面处理
Electrical discharge machining (EDM)	电火花加工
Bench work	钳加工
Assembly	装配
Gear machining	齿加工
Sheet metal forming	钣金成型
Welding	焊接
Wire EDM	线切割
Laser marking	激光打标
Inspection	检验
Stock cutting	下料

Quickstart

1. Try the online demo

Use the LevelField demo to explore the drawing-to-process-planning workflow without a local deployment.

2. Download the model

pip install -U huggingface_hub
hf download level-ai/LevelField --local-dir ./LevelField

3. Start a vLLM server

The release documents these deployment requirements:

• vLLM 0.19.1 or later.
• At least one NVIDIA H100 GPU.
• NVIDIA CUDA 12.9, as reported in the tested environment.

GPU memory capacity, concurrency, and image size were not specified in the original deployment notes. Memory requirements depend on workload and serving settings; the approximately 20 GB weight size is not the total GPU memory requirement.

In a compatible CUDA environment, install vLLM and launch the server:

pip install 'vllm>=0.19.1'

CUDA_VISIBLE_DEVICES=0 \
OMP_NUM_THREADS=1 \
vllm serve ./LevelField \
    --host 127.0.0.1 \
    --port 18124 \
    --served-model-name LevelField \
    --reasoning-parser qwen3 \
    --chat-template-content-format openai \
    --trust-remote-code \
    --max-model-len 32768 \
    --gpu-memory-utilization 0.9 \
    --mm-processor-cache-type shm

This configuration uses a 32,768-token serving limit and binds the endpoint to the local machine. The serving limit is a deployment setting, not a claim about the model's maximum supported context length.

4. Submit a drawing

Install the client:

pip install openai

Save a legible drawing as test.jpg in your working directory, then run the following request example. The Chinese prompt asks the model to generate a process sequence using the 18 available process types.

import base64
from openai import OpenAI

client = OpenAI(api_key="EMPTY", base_url="http://localhost:18124/v1")
model = "LevelField"


def encode_image_base64(image_path):
    with open(image_path, "rb") as f:
        return base64.b64encode(f.read()).decode("utf-8")


image_path = "test.jpg"
image_base64 = encode_image_base64(image_path)

response = client.chat.completions.create(
    model=model,
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "text",
                    "text": (
                        "您是一位工艺工程师。现在给您加工图纸<image>。"
                        "请根据可供选择的工艺`车削,铣削,钻削,磨削,铸造,锻造,"
                        "热处理,表面处理,电火花加工,钳加工,装配,齿加工,"
                        "钣金成型,焊接,线切割,激光打标,检验,下料`，"
                        "给出它的工艺序列。"
                    ),
                },
                {
                    "type": "image_url",
                    "image_url": {
                        "url": f"data:image/jpeg;base64,{image_base64}"
                    },
                },
            ],
        }
    ],
    temperature=0.01,
    max_tokens=2000,
    extra_body={
        "chat_template_kwargs": {"enable_thinking": True}
    },
)

message = response.choices[0].message
if hasattr(message, "reasoning") and message.reasoning:
    print("【思考过程】")
    print(message.reasoning)
    print()
print("【最终回答】")
print(message.content)

This example enables thinking through chat_template_kwargs and prints the reasoning field when returned, followed by the final answer. Use it with the server configuration above, including --reasoning-parser qwen3.

For a PNG drawing, change the file path and the data URL MIME type to image/png. The image payload follows the vLLM multimodal input format.

If generation reaches the 2,000-token output limit before producing a complete answer, increase max_tokens within the available context capacity.

Recommended workflow

1. Supply a legible drawing with all relevant views and annotations.
2. Check the extracted requirements against the drawing, especially units, tolerances, materials, and surface finish.
3. Provide production constraints such as quantity, stock form, available machines, tooling, and inspection capabilities.
4. Request a process route using the reviewed requirements and constraints.
5. Have a qualified process engineer validate the route before production use.

Training background

Level AI reports fine-tuning and reinforcement training using 100,000 manufacturing drawings and process sheets from 16 factories. Data construction and result validation involved the Level AI engineering team and 36 process engineers with more than 20 years of hands-on experience.

The release does not specify the data split, training hyperparameters, reinforcement method, or dataset access and licensing details.

Reported results

Evaluation setting	Reported result
Internal evaluation	More than 95% alignment between model process reasoning and senior process engineers' decisions
Pilot customer usage	Approximately 60% reduction in process-route drafting time

These are developer-reported results. The release does not provide evaluation sample sizes, scoring criteria, comparison baselines, or uncertainty estimates. The alignment figure should not be interpreted as a general accuracy score, and the time reduction may vary across production settings.

Limitations and intended use

LevelField is intended to assist process engineers with interpretation and planning. Generated routes are proposals that require engineering review.

• Check small annotations, dense drawings, and ambiguous geometry carefully; extracted details may be incomplete or incorrect.
• Missing information can lead to unsupported assumptions. Ask the model to identify uncertainty explicitly.
• A drawing alone does not establish shop-floor feasibility. Validate machine capability, fixtures, tooling, material condition, and inspection requirements.
• Performance across different industries, drawing conventions, materials, and languages has not been separately documented in this release.
• The release does not establish validated CNC program generation or autonomous production control.

License and feedback

LevelField is released under Apache 2.0. Review the repository's license terms before redistribution or commercial integration.

For questions or reproducible issues, open a discussion in the Hugging Face community tab. Include the serving version, configuration, prompt, and a shareable example where possible.