PickQwen-2.5-1B-RLCD: Parallel Decoding Boosts JSON Extraction Speed 7x on M4 Max

Maximizes structured data extraction speed using the MLX framework on Apple Silicon. Instead of generating tokens one by one as in traditional methods, it broadcasts the KV-Cache in parallel to evaluate multiple fields simultaneously. This achieves up to a 7x reduction in latency on the M4 Max.

It guarantees 100% valid syntax by verifying only the candidate sets for each field in the JSON schema via logit slicing. It supports complex tasks such as high-cardinality classification with up to 255 options or enterprise ticket triage with 28 fields. It returns calibrated confidence scores for all fields, making it easy to verify the accuracy of the results.

Based on the Qwen-2.5-1B model, it enables lightweight inference in local environments. A web interface allows visual comparison of performance differences between parallel decoding and standard autoregressive generation. It is suitable for classification and routing tasks requiring fast responses, such as financial fraud detection or code security audits.