level-ai/LevelField
LevelField is a 9-billion parameter multimodal model developed by Level AI, based on Qwen3.5-9B. It specializes in interpreting 2D manufacturing drawings and generating non-standard machining process plans, including operation sequences and optimization suggestions. The model supports both Chinese and English, excelling at extracting details like dimensions, tolerances, and materials from drawings to assist process engineers. It is designed for manufacturing process planning across 18 different process types.
Loading preview...
LevelField: Specialized Multimodal AI for Manufacturing Process Planning
LevelField, developed by Level AI, is a 9-billion parameter multimodal model built upon the Qwen3.5-9B architecture. It is specifically designed to assist process engineers in non-standard machining process planning by interpreting 2D manufacturing drawings and drafting optimized machining routes. The model supports both Chinese and English.
Key Capabilities
- Drawing Interpretation: Extracts critical information from PDF and image documents, including part names, drawing numbers, materials, dimensions, tolerances, surface roughness, and technical requirements.
- Process Planning: Proposes operation sequences, identifies unsuitable processes, and suggests improvements based on reviewed machining requirements.
- Extensive Process Coverage: Supports 18 distinct manufacturing process types, ranging from turning and milling to casting, forging, and various surface treatments.
Training and Performance
LevelField was fine-tuned and reinforced using 100,000 manufacturing drawings and process sheets from 16 factories, with validation from 36 experienced process engineers. Internal evaluations report over 95% alignment between the model's process reasoning and senior engineers' decisions, and pilot customer usage showed approximately a 60% reduction in process-route drafting time.
Intended Use and Limitations
This model is intended as an assistive tool for process engineers, providing proposals that require human review. Users should carefully check extracted details, especially for small annotations or ambiguous geometry. The model does not establish validated CNC program generation or autonomous production control. It is licensed under Apache 2.0.