Qwen-2.5-1B-RLCD Runs 7x Faster on Apple Silicon

An independent developer has launched Qwen-2.5-1B-RLCD, an Apache 2.0-licensed inference engine designed to accelerate structured classification and data extraction on Apple Silicon. By utilizing MLX to score multiple JSON fields in parallel, the engine achieves a 5.6x to 7x speedup in JSON generation compared to traditional sequential methods. The tool runs locally on M1 through M4 Macs and requires no fine-tuning or retraining, wrapping the existing mlx-community/Qwen2.5-1.5B-Instruct-4bit model.

Traditional JSON generation relies on autoregressive decoding, where a model generates output token by token. This sequential process requires a new forward pass for every character, punctuation mark, or field name, causing latency to spike as schemas grow. Qwen-2.5-1B-RLCD bypasses this bottleneck by treating each field as a classification problem. It evaluates all schema fields simultaneously against a single broadcast key-value cache, scoring permitted values from a predefined set of up to 255 enum choices per field. The engine then programmatically assembles the final JSON object.

This parallel approach yields massive performance gains. In a benchmark involving a 28-field enterprise triage task, processing time plummeted from 1,900 milliseconds and 312 forward passes down to just 270 milliseconds in a single pass. Despite the speed, the engine maintains 100% schema validity with calibrated per-field confidence probabilities across all tested scenarios. For practitioners, this means local edge deployments on Mac hardware can now handle complex, structured data extraction tasks with near-instantaneous response times, making real-time local classification highly practical.