VittoriaLanzo/Ohmatic-Qwen3-8B

Ohmatic-Qwen3-8B: Self-Correcting Circuit Schematic Generation

Ohmatic-Qwen3-8B is an 8 billion parameter model, fully merged from Qwen3-8B, engineered for generating electronic circuit schematics from natural language. Its core innovation lies in a closed verification loop that enables self-correction of designs.

Key Capabilities

• Forward Generation: Translates natural language circuit descriptions into structured schematics (components, values, nets).
• ERC Verification: Integrates with an Electrical Rule Checker to identify design flaws like shorts, floating nets, or polarity errors.
• Self-Correction: Uniquely trained to receive ERC feedback and emit a repaired schematic, improving designs rather than simply re-rolling them. This explicit training for correction turns is a key differentiator.

Training and Performance

The model is trained to both produce circuits and repair its own designs using only ERC-verified examples. It achieves a 53.1% first-pass validity on held-out ERC evaluations, with the correction loop significantly recovering further failures. The training data, recipe, and ERC engine are proprietary. The model is released as fully merged weights, requiring no PEFT/adapter.

Usage and Licensing

Ohmatic-Qwen3-8B can be loaded using transformers for bfloat16 inference or via llama.cpp with GGUF quantizations (8-bit and 4-bit). It operates under the Ohmatic Source-Available License 1.0 (Ohmatic-SAL-1.0), which permits most uses but restricts 'Competing Uses' for 10 years before converting to Apache-2.0.