mizorewww/laya-mlx is an independent port that runs Laya's type-based decision models on Apple Silicon with MLX. The latest confirmed release is v0.5.0. The provided materials do not list the changes in this version.
This runtime makes decisions on given questions instead of continuing sentences. `choice` returns the probability of each option, `score` returns the expected score according to the scoring criteria, and `noul` returns the probability that a proposition is true. There is no need to generate JSON with a bidirectional encoder and parse it afterwards. It provides English and multilingual typed-decisions checkpoints along with a Python API, CLI, and Snake demo, while RLCD training and fine-tuning are done in the upstream project. It runs in MLX without PyTorch or cloud APIs, and the public weights and judgment probability outputs make it a setup that can be used for local agent routing.
The short decision latency stated in the repository description is 7–14 ms on an M3 Max. It reports an immediate-execution control group and a paired comparison under the same execution conditions for the compiled path and the prefix-reuse path, but there is no basis for extending these numbers to the performance of other hardware. Since the API benchmark for a single question must be distinguished from real game-loop latency, the stated latency can only be taken as a reference within that device and measurement scope.