Two efficient-inference releases hit Hugging Face trending on 2026-09-16: Parallel Constrained Decoding MLX engine (#30) and Prism ML's ternary Ternary-Bonsai-2-27B (#29)

Two Apple Silicon–relevant efficient-inference items appeared on Hugging Face trending: an MLX-based Parallel Constrained Decoding engine reporting 5.6x-7.0x latency speedups for schema-valid JSON (trending #30) and Prism ML's Apache 2.0 Ternary-Bonsai-2-27B,…

On 2026-09-16, two model/engine releases appeared on the Hugging Face trending list. At #30, the repository harshatheg/Qwen-2.5-1B-RLCD — whose content actually describes Parallel Constrained Decoding, an MLX-based inference engine for structured extraction and classification on Apple Silicon Macs — benchmarked with mlx-community/Qwen2.5-1.5B-Instruct-4bit on an M4 Max reports 5.6x-7.0x latency reductions (e.g., 1,900 ms down to 270 ms for a 28-field support triage task) with 100% syntactic validity and calibrated field-level probabilities. It prefills a single KV-cache, broadcasts it across all schema fields for simultaneous evaluation, and slices logits to valid candidate tokens for enum fields with up to 255 choices; it requires Apple Silicon (M1-M4) and macOS 14+. At #29, Prism ML published Ternary-Bonsai-2-27B (repo prism-ml/Ternary-Bonsai-2-27B-gguf), an Apache 2.0 open-weights model derived from Qwen3.8-27B (27.36B parameters, hybrid attention ~75% linear, 262K-token context). Its weights are end-to-end ternary {-1, 0, +1} with g128 FP16 scales at 1.72 bits/weight, shrinking the model from ~54 GB FP16 to 5.95 GB (PTQ1_0) or 7.21 GB (PQ2_0) while reportedly retaining 98.2% of FP16 quality (84.78 average across 14 thinking-mode benchmarks). It ships as GGUF for llama.cpp (CUDA, Metal, CPU) with custom ternary kernels, plus an MLX 2-bit build for Apple Silicon and an optional Q8_0 vision tower pack, alongside a whitepaper, demo repo, and forked llama.cpp/MLX runtimes.

• 2026-09-16: harshatheg/Qwen-2.5-1B-RLCD appeared at #30 on Hugging Face trending; its content describes Parallel Constrained Decoding, an MLX-based inference engine for structured extraction and classification on Apple Silicon Macs.
• Parallel Constrained Decoding benchmarked with mlx-community/Qwen2.5-1.5B-Instruct-4bit on an M4 Max reports 5.6x-7.0x latency reductions, e.g., 270 ms versus 1,900 ms for a 28-field support triage schema.
• The engine prefills a single KV-cache and broadcasts it across all schema fields, evaluating them simultaneously instead of token-by-token autoregressive generation.
• Sub-vocabulary logit slicing over candidate sets yields 100% valid JSON with calibrated softmax field-level confidence scores; enum fields support up to 255 categorical choices.
• Parallel Constrained Decoding requires Apple Silicon (M1-M4) and macOS 14+.
• 2026-09-16: prism-ml/Ternary-Bonsai-2-27B-gguf appeared at #29 on Hugging Face trending; Prism ML released Ternary-Bonsai-2-27B, an Apache 2.0 open-weights model derived from Qwen3.8-27B.
• Ternary-Bonsai-2-27B has 27.36B parameters, hybrid attention ~75% linear, and 262K-token context.
• Weights are end-to-end ternary {-1, 0, +1} with g128 FP16 scales at 1.72 bits/weight, shrinking the model from ~54 GB FP16 to 5.95 GB (PTQ1_0) or 7.21 GB (PQ2_0).