Laya-MLX: Native MLX Runtime for Laya Typed Decision Models

• Fact Laya-MLX runs natively on Apple Silicon and achieves 7–14 ms latency for short decisions on M3 Max.[1]AINative MLX runtime for Laya typed decision models — 7–14 ms short decisions on M3 Max.
• Fact The median end‑to‑end latency for a short English typed decision is 13.4 ms, dropping to 7.4 ms with a multilingual checkpoint.[1]AI13.4 ms median end-to-end for a short English typed decision. 7.4 ms with the multilingual checkpoint.
• Fact Inference is fully local, requiring no PyTorch, Transformers runtime, or cloud API.[1]AILocal MLX inference, with no PyTorch, Transformers runtime, or cloud API.
• Fact The runtime supports typed decisions such as choice probabilities, rubric score probabilities, and proposition truth probabilities.[1]AI- choice: probabilities over named options. - score: probabilities over ordered rubric levels and their expected score. - noul: P(true) for a proposition.
• Fact All three checkpoints matched the upstream selected answer on 63/63 validation questions in both FP32 and FP16, with no active‑memory growth over 100 repeated calls.[1]AIPort fidelity: all three checkpoints matched the upstream selected answer on 63/63 validation questions in both FP32 and FP16 — 378/378 comparisons. Each configuration also passed 100 repeated finite, deterministic calls with zero measured active-memory growth.
• Fact The GitHub repository mizorewww/laya-mlx was created on 19 Sept 2026 and had 6,715 stars on 2 Oct 2026.[1]
• Fact Its description reads: “Native MLX runtime for Laya typed decision models — 7–14 ms short decisions on M3 Max. No text generation, PyTorch, or cloud API.”[1]
• Fact That is 6,715 stars in 13 days, an average of 507.6 per day.[1]

• Analysis Integrate Laya-MLX into macOS applications to perform low‑latency typed decisions locally, leveraging the provided CLI demo or embedding the MLX runtime in Python code.AI
• Analysis Automatically detected from public activity data. Not yet reviewed by an analyst.

• Fact Laya-MLX runs natively on Apple Silicon and achieves 7–14 ms latency for short decisions on M3 Max.[1]AINative MLX runtime for Laya typed decision models — 7–14 ms short decisions on M3 Max.
• Fact The median end‑to‑end latency for a short English typed decision is 13.4 ms, dropping to 7.4 ms with a multilingual checkpoint.[1]AI13.4 ms median end-to-end for a short English typed decision. 7.4 ms with the multilingual checkpoint.
• Fact Inference is fully local, requiring no PyTorch, Transformers runtime, or cloud API.[1]AILocal MLX inference, with no PyTorch, Transformers runtime, or cloud API.
• Fact The runtime supports typed decisions such as choice probabilities, rubric score probabilities, and proposition truth probabilities.[1]AI- choice: probabilities over named options. - score: probabilities over ordered rubric levels and their expected score. - noul: P(true) for a proposition.
• Fact All three checkpoints matched the upstream selected answer on 63/63 validation questions in both FP32 and FP16, with no active‑memory growth over 100 repeated calls.[1]AIPort fidelity: all three checkpoints matched the upstream selected answer on 63/63 validation questions in both FP32 and FP16 — 378/378 comparisons. Each configuration also passed 100 repeated finite, deterministic calls with zero measured active-memory growth.

Architecture

Insufficient evidenceNo architecture has been documented for this signal.

Insufficient evidenceDeveloper implications have not been assessed yet.

Insufficient evidenceNo adoption indicators recorded.

Timeline

1. 2 Oct 2026

   First seenFirst seen by the Astrava pipeline[1]

• Fact Performance measurements were conducted on a single M3 Max machine; older macOS versions were not tested.[1]AIOlder supported macOS versions were not tested on this machine.
• Analysis Generated from activity metrics only; verify the linked sources before acting on it.

Insufficient evidenceNo competing approaches documented yet.

Insufficient evidenceNo independent projects or organisations are documented as building on this yet.