What is the skycmd project?

A project that trains tiny System One decision models from scratch in Rust with MLX on a MacBook and plays them live in the browser through WebAssembly, with model cards on Hugging Face.

How small are the models?

The highlighted example, s3-pico, has 445k parameters and a 251 KB ternary weight file; it plays four games, including a Doom-like one, at oracle level in the browser at about 50 microseconds per frame.

Which models are published?

Six: s5-lite, s4-micro, s3-pico, s2-nano, s1-pico and s2-nano-doom, each with a Hugging Face card and a browser demo, published with the scripting in scripts/hf-publish.py.

How does a System One decision model work here?

It receives a state plus typed questions, choice, score and noul for probabilistic yes/no, and returns a calibrated probability for every allowed option in one forward pass, without generating text.

Yes, the project links a tutorial page alongside play, live, results, install and bench pages at maxgfr.github.io/skycmd; the digest does not include the tutorial text itself.

What license and technology stack?

Apache-2.0, written in Rust, with MLX training on Apple Silicon, GGUF export and WebAssembly SIMD inference; its topics include llama-cpp, modernbert, tiny-ml and drone.

What is missing from the digest?

The captured README stops inside the model table, so per-model benchmarks, the closed-loop sky-arena play results and training hardware details are not available; GitHub shows 0 stars and an Apache-2.0 license.

Jev directory

CommunityJev Models & Reimplementations

Skycmd

Tiny System One decision models trained from scratch in Rust with MLX on a MacBook and played live in the browser via WebAssembly: s3-pico packs 445k parameters into a 251 KB ternary file and decides in about 50 microseconds per frame.

See highlights Open github.com

Stars: 0
Language: Rust
License: APACHE-2.0
Reading time: 2 min
Published by: Community
Author: maxgfr
Last verified: Oct 2026

maxgfr/skycmd

User repository on GitHub

View on GitHub

Tiny System One decision models trained from scratch in Rust with MLX on a Mac. s3-pico: 445k params, 251 KB ternary, plays four games (incl. a Doom-like) at oracle level live in the browser via WebAssembly, ~50 µs per frame.

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

Tagscommunity github rust wasm

Works forJev for Classification Jev for Scoring

Used inJev for Gaming

Languages: Rust
Platforms: Browser
Capabilities: Open source

Highlights

Tiny System One decision models trained from scratch in Rust with MLX on a MacBook, then played live in the browser through WebAssembly.
s3-pico: 445k parameters, a 251 KB ternary model that plays four games, including a Doom-like one, at oracle level, about 50 microseconds per frame.
Six published models from s1-pico to s5-lite plus s2-nano-doom, each with a Hugging Face card and browser demo; publishing uses scripts/hf-publish.py.
Decision models take a state and typed questions (choice, score, noul) and return calibrated probabilities in one forward pass, with no text generation.
Apache-2.0 Rust project with play, live, results, install, tutorial and bench pages; 0 stars, created 2026-10-11, pushed 2026-10-14.

Watch out

README reviewed 2026-10-17; the captured excerpt cuts off inside the model comparison table, so the full benchmark numbers, training recipe details and hardware requirements beyond a MacBook are not available in the digest.

Questions about this entry

Short answers drawn from the linked project or page.

From the Jev guides

Original write-ups on what this entry is for.

Beyond text and boxes: multimodal decisions you can run yourself

Oct 16, 2026

Jev-Omni's merged classifier scores text, image, audio, and video options; VideoJev and Laya Vision show what open multimodal decision models accept, what they return, and which numbers are the authors' own.

Jev Models & Reimplementations#multimodal#open-models#vision

Read article

Jev 1.14: Images, Box, and the Danger Zone

Oct 15, 2026

Jev Day shipped a routine model bump and a deliberate widening of the API: images on both models, a box question type on 1.14 only, and 262k context, all behind danger-zone headers.

Jev Models & Reimplementations#release#vision#danger-zone

Read article

The Open Decision-Model Roundup

Oct 14, 2026

Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.

Jev Models & Reimplementations#open-models#training#local

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai

Skycmd

Highlights

Questions about this entry

What is the skycmd project?

How small are the models?

Which models are published?

How does a System One decision model work here?

Is there a tutorial?

What license and technology stack?

What is missing from the digest?

More Jev resources like this

From the Jev guides

From the Jev AI model community