s3-pico

A 251 KB decision model that plays four games in your browser. 445,345 ternary parameters: one weight-shared state block applied up to 4 times, with an exit after each pass. It reads the game state as text, answers typed questions (choice, score, noul) in one forward pass each, and a coded safety reflex keeps a veto.

• Demo: huggingface.co/spaces/maxgfr/s3-pico (watch it play, or take the controls)
• Code: github.com/maxgfr/skycmd (100 % Rust)
• Story: chapter 21

Results (closed loop, sky-arena, 30 episodes per game, seed 7)

game	result
Flappy Drone	200 / 200 pipes, 0 crashes
Rescue Run	100 %
Obstacle Dodge	87 % (= oracle), 14 reflex vetoes
Doom-like	100 % of levels, kills = oracle

It uses 1.05 to 1.73 of its 4 loops on average (exit threshold calibrated by playing). Latency per decision: 27.5 µs native, 49.6 µs in WASM (Apple M5).

Architecture

The s3 format (skycmd-s2 GGUF tag): question and options form a static stream encoded once and cached; the state is one token per key value pair, the value added to the key's embedding; one weight-shared state block looped 4 times with an exit after each pass; options never attend to each other; ternary weights trained with QAT.

Usage

Runs in the skycmd engine only (browser, Node, Python, CLI, server):

cargo install --git https://github.com/maxgfr/skycmd sky-serve
skycmd decide --model s3-pico.gguf --request req.json --pretty

Request format and bindings: install guide. The games are civil drone scenarios and a Doom-like shooter drawn with Freedoom art; the model only advises and coded rules always keep a veto. This is a demo and a tutorial, not flight software.

Training

From scratch in Rust (mlx-rs) on Apple Silicon: numeric masked-language pretraining, decision training on soft targets with deep supervision over the exits, ternary QAT, loop dropout, DAgger rounds on the games and closed-loop exit calibration. Data: typed-decision sets, BoolQ, MNLI, ARC, CommonsenseQA and Banking77 converted into typed questions, SimpleStories, TinyStories and FineWeb for pretraining, and game states labelled by exact oracles. Part of the training data was generated with GLM. Licenses: DATA_LICENSES.md.

Files

• s3-pico.gguf: ternary weights, tokenizer, temperatures and exit thresholds.