Laya MPS
Runs Laya's typed-decisions checkpoint on Mac GPUs via PyTorch MPS, with a low-memory mode down to about 0.74 GiB.
Overview
Laya MPS runs Jev-style typed decisions locally on Apple Silicon through PyTorch's Metal Performance Shaders backend, using Laya's English typed-decisions checkpoint, which the README describes as specialised in customer service, invoices, security incidents and agent traces rather than a general language model. A script installs dependencies, downloads the roughly 843 MB model and starts a local server with a demo page that can play a game or run a 60-second latency benchmark.
Its focus is memory. Measured on an M5 Pro with 24 GiB, the default reduced setting peaks at 2.11 GiB with 32 ms median decision latency, full uses 2.30 GiB at the same speed, and minimal streams layers and embeddings from disk to peak at 0.74 GiB with 171 ms latency. All three settings matched exactly on 260 of 260 decisions including probabilities, because computation stays in FP32. It suits Mac users who want offline decisions on memory-constrained machines.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which option fits the input? | — |
| Score | How should the input be scored? | — |
| Yes / no | Is this statement true? | — |
How it works
- Clone the repo and run ./scripts/serve.sh
- First run installs dependencies and downloads the model
- Choose a memory setting
- Send decisions to the local server or use the demo
Strengths
- Very low memory option
- Identical answers across memory modes
- Offline after first download
Caveats
- English, domain-specialised checkpoint
- Minimal mode is slower
- Apple Silicon and macOS 14+ only
Getting started
git clone https://github.com/afshinm/laya-mps.git && cd laya-mps && ./scripts/serve.sh, then open http://127.0.0.1:8000/demo/.At a glance
- Type
- Model
- Status
- beta
- Language
- Python
- Runtime
- PyTorch MPS server on macOS
- Has code
- yes
Use cases and integrations
How curated lists describe it
Runs Laya's typed-decision checkpoints on Apple Silicon through Metal Performance Shaders, with a lower-memory mode.
MrJev/awesome-jev
Frequently asked questions
How little memory can it use?
About 0.74 GiB in minimal mode, at 171 ms per decision.
Do answers change with memory mode?
No, all modes matched on 260 of 260 decisions.
Is the model general-purpose?
No, the README says it specialises in customer service, invoices, security incidents and agent traces.
Related projects
Docker service exposing local Laya, Von, a CLM head or hosted Jev through one TypeSafe-compatible HTTP and MCP endpoint.
MCP serverbetaPythonConverts open Jev reproductions (Laya, Kev, NanoJev, PlayJev) to INT8 ONNX for offline CPU decisions at about 15.6 ms per question.
ServicebetaPythonNative Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.
ProjectbetaGoGo harness for System One models: hosted Jev, open Kev in-process via llama.cpp, and Cloudflare Clef, with typed helpers and a CLI.
ServicebetaGoGo gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.
ServicebetaGoRuns Intern-Decision-0.8B on Apple Silicon behind a local /v1/systemone endpoint that accepts screenshots for agent decisions.
ServicebetaPython