Is it actually faster than laya serve?

On the measured base machine it answered a three-question request in about 17ms of inference and beat stock serving by multiples per request, with a smaller edge over tuned Python. Answers matched the upstream reference on nearly all of 1,500 checks, with both misses near ties.

Do I need to rewrite my Jev client?

No. The server speaks the same endpoint, limits, error codes, and detail strings, so existing Jev and Laya clients point at its port unchanged. It never downloads models itself and reads weights from the local cache.

Which confidence number gates actions?

Use answer_confidence for whether to trust an answer, not the entropy-based confidence that only measures spread. A separate action probability estimates acting versus escalating; never compare the two against one threshold.

What happens under load?

One request runs while up to 16 wait; a full queue returns 503 with Retry-After immediately, so clients must retry. More clients do not add throughput because a single request already fills the GPU.

How do I install sys1rust?

On an Apple silicon Mac with macOS 14 or later, run the install.sh one-liner or paste a setup line into a coding agent. It installs into ~/.local/share/sys1rust without sudo and links ~/.local/bin/sys1rust.

How big is the model download?

The first sys1rust serve downloads the typed-decisions model, 846 MB, into the Hugging Face cache. sys1rust pull downloads a model ahead of time and sys1rust models lists what is downloaded.

Does it give the same answers as the reference?

Yes, within a stated gate. On the 1,500-answer correctness workload, 1,498 agree with the upstream PyTorch fp32 reference, 99.9% against a 99% gate, and the two that differ are near ties.

CommunityJev Models & Reimplementations

Sys1rust

Rust server running Laya decision models locally on Apple silicon GPUs over MLX with no Python at runtime, speaking the same /v1/systemone API as laya serve.

See highlights

Open github.com

Stars: 3
Language: Rust
License: APACHE-2.0
Published by: Community
Author: krishhgg
Last verified: Oct 2026

krishhgg/sys1rust

User repository on GitHub

View on GitHub

Laya's System 1 decisions in 15.5 ms on an M5 Mac, up to 4.1x faster than laya serve and 1.25x faster than Python MLX. Same /v1/systemone API, written in Rust with custom Metal kernels, no Python.

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

sys1rust runs Laya's System 1 models on your Mac's GPU, in Rust with its own Metal kernels and no Python. Send it some text and a few questions, and it answers each one with a choice, a score or a yes probability. Its server, sys1rust serve , speaks the /v1/systemone API of laya serve , so Jev and Laya clients send it the same requests. Paste this line into Claude Code, Codex, Cursor or another coding agent:

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Rust

Highlights

Three-question request answers in about 17ms of inference on the measured base machine GPU.
Matches laya serve key checks, limits, error codes, and detail strings, queuing up to 16 slots.
Batches a request's questions as one row per question in fp16, skipping padding work unchanged.
Reads weights only from the local cache with typed-decisions, multilingual, and english options.

Watch out

Apple-silicon Mac build with Rust, CMake, Xcode tools, and a venv-held MLX path; one request fills the GPU so concurrency queues rather than scaling. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

Ollaya

1.3k

A local daemon that pulls open decision models and serves them on TypeSafe's /v1/systemone, so a Jev client switches hosts with TYPESAFE_BASE_URL.

Jev Models & Reimplementations#community#rust#open-models

Community

ollaya-dev/ollayaRust

system-one-foundation-models

Native Swift 6 package that evaluates @Generable structs and enums through System One decision models in 15-150 ms, with on-device Laya, self-hosted laya-serve, and TypeSafe Jev cloud backends.