TL;DR

On September 16, 2026, Harsha Gundala released Qwen-2.5-1B-RLCD, a small MLX engine for typed JSON on Apple Silicon. It prefills the context and schema once, scores each enum or boolean field over a masked candidate set, and assembles the object in code. The model card reports about 5.6x to 7x lower latency than autoregressive structured generation on an M4 Max, with schema-valid output by construction.

Why it matters

Routing, triage, and classification rarely need a paragraph. They need a choice from a closed list, and a confidence beside it. Ordinary structured generation still walks the braces, keys, and spaces one token at a time, so latency grows with the JSON rather than with the decision.

Parallel constrained decoding is built for that narrower job. One KV-cache is shared across fields. Each field is a score over its allowed labels. The program writes the JSON from those labels.

A Decisions-style API is the hosted form of the same job: a service contract for categorical judgments, with a bounded answer and a confidence, rather than prose. Gundala's release is the decoding method on a small local model. Use it for on-device, multi-field schema fill when every field is a closed choice. Use a Decisions-style API when that judgment should be a hosted call. This page does not describe an OpenAI endpoint, request format, or benchmark.

Key takeaways

• Prefill the document and the field descriptions once, then reuse that KV-cache for every field.
• Enums and booleans are scored over their candidate tokens. The rest of the vocabulary is masked, and the card reports a softmax confidence on that slice.
• The program assembles the JSON from those choices, so broken syntax is not a token the model has to avoid.
• Reported timings on an M4 Max use mlx-community/Qwen2.5-1.5B-Instruct-4bit, even though the release is named Qwen-2.5-1B-RLCD: 5.6x for 4-field cases and a 255-choice tariff field, and 7x for a 28-field triage schema. Those figures are the author's benchmarks.
• Free-form text is out of scope. A field that has to be written still needs autoregressive decoding.
• The public demo space is paused. The model card and the announcement post are the durable sources.
• The Decisions comparison is the product shape, a closed choice plus a confidence. It is not a claim of shared code or a shared interface.