Should you choose Laya-MLX?
Laya-MLX runs open-weight typed decisions natively on Apple Silicon. Use it when your application needs a probability over named options, an expected rubric level or P(true), and you can test the result on the Mac that will run it. Its output format removes the need to parse generated prose, but it does not establish whether an answer is correct.
This review separates the recorded v0.3.0 checkout from the current documentation. Our run tested commit ca5940a, not the later implementation described by the current main-branch README.
What our sandbox established
Installation took 17 seconds; the build took 5 seconds. Our run took place on October 6, 2026, at commit ca5940a in a Debian Bookworm container. The recorded footprint was 51 packages and 150 MB before any checkpoint download.
The harness stores each command's outcome, exit code, elapsed time and log tail. Here, the test command returned exit 4 after 3 seconds. Its tail identifies an import failure in tests/conftest.py: import mlx.core as mx raised ModuleNotFoundError: No module named 'mlx'. The stored test-count field is null, so there is no passing-test total to report.
That result establishes a limit of this sandbox run. It does not establish a defect on supported Apple hardware, and the zero exit codes for install and build do not demonstrate a working model or CLI. We did not test checkpoint loading, inference accuracy, latency, memory under inference, offline operation or service deployment.
The MrKeyoor testing methodology explains how command outcomes reach the review. That supplied page describes the publisher's current procedure; it does not establish that its current version governed this older run.
Pip-audit recorded 0 advisories, with unknown scan coverage. This is an advisory result, not a security assessment.
Check the version before rejecting a feature
The README pinned to our checkout describes inference and conversion, with training and fine-tuning left upstream. It explicitly excludes upstream batch, long-document, hooks and server APIs. The v0.3.0 release also calls its upstream compatibility selective and says real-checkpoint accuracy and long-running memory benchmarks were not rerun for that release.
Those exclusions should not be presented as current project limitations. The current README now documents batch prediction, long-text scanning, confidence review flags, CLI, HTTP and MCP interfaces, and integration extras. It says host APIs follow upstream v0.4.17. These are documented capabilities of the current source, not capabilities verified by our October sandbox run. The exact version assessed by our run remains v0.3.0.
Evaluate the decision, not just its confidence
The pinned README warns that choice options share a token budget. A large label set can leave too few tokens to distinguish options. Its optional shortlisting helper reduces the set before prediction, so the returned probabilities cover only the retained labels. If excluding the correct label would be costly, evaluate the shortlist separately from the final decision.
The project reports checkpoint agreement with upstream on finite validation fixtures. Its benchmark documentation distinguishes that fidelity from task accuracy. Published Mac latency figures also exclude model loading and depend on workload, precision and hardware; they are not our measurements. The current benchmark page explicitly separates later qualification work from historical experiments and says performance was not re-measured for release 0.6.0.
A useful next step is to pin the runtime and checkpoint, download the weights on your supported Mac, and evaluate representative labeled inputs. Include long states, ambiguous requests and your largest choice sets. Measure model loading separately from steady-state calls, and decide how your application will handle uncertain or wrong answers.
Which alternative fits the job?
Upstream Laya provides fine-tuning workflows and CPU or GPU PyTorch installation guidance. Consider it as a replacement runtime when your platform or training requirements extend beyond this MLX port. Its documentation includes platform-specific setup caveats, so a broader platform path still needs validation in your environment. The existing Laya review covers that implementation separately.
MLX LM generates text and fine-tunes large language models on Apple silicon. Choose it for prose or chat generation. In an application that needs both generated text and constrained decisions, it can complement Laya-MLX; it is not a drop-in implementation of Laya's typed-decision interface.
How the review scores work
Assessed scores use a 1-to-5 rubric: 1 means substantial documented obstacles, 2 means major gaps, 3 means usable with material constraints, 4 means strong evidence with limited gaps, and 5 means unusually complete evidence for that dimension. Setup considers prerequisites and observed command outcomes; documentation considers actionable coverage and stated limits; maturity considers release scope and validation boundaries. Community requires evidence about participation and support responsiveness. Stars and an empty issue queue alone do not establish it.

