s2-nano

A 702 KB ternary decision model with a cached question stream and early exits. 1,532,033 parameters. It answers typed questions about a game state in one forward pass each and stops at the first exit confident enough.

• Demo: huggingface.co/spaces/maxgfr/s2-nano
• Code: github.com/maxgfr/skycmd (100 % Rust)
• Story: chapter 19, the s2 architecture

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), 25 reflex vetoes

Latency per decision: 273.0 µs native, 409.1 µs in WASM (Apple M5). At the exit thresholds calibrated on validation questions it crashed in every Flappy episode; at τ = 0.95 (shipped) it plays at the oracle's level using 3.2 to 3.4 of its 4 layers. Calibrate early exits on closed-loop states.

Architecture

The s2 format (skycmd-s2): a static stream (question + options) encoded once and cached, a state stream encoded once per frame, options that never attend to each other, one [NUM] token per number with continuous value features, early exits with deep supervision, ternary weights trained with QAT. Spec: SPEC-S2.md.

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 s2-nano.gguf --request req.json --pretty

Request format: install guide.

Training

From scratch in Rust (mlx-rs) on Apple Silicon: numeric masked-language pretraining, decision training with deep supervision over the exits, ternary QAT, Muon and AdamW with a WSD schedule, DAgger rounds on the games. 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

• s2-nano.gguf: ternary weights, tokenizer, temperatures and exit thresholds.