Moonshine
Speech becomes text on your device.
LESS FRICTION.
MORE COMMAND.
Speak an order. Ask for an assessment. Stay with the battle.
LFM2.5 · 350M
A fine-tuned model produces a constrained command.
The simulation
Game rules and policy authorize the actual action.
Kokoro
Grounded feedback becomes a spoken response.
YOUR INTENT.
THE GAME’S RULES.
The model does not invent resources, bypass research, or change the battle directly. It translates language into commands the game can inspect and execute. An analysis request should provide information; it should not spend your materials.
Typed plans, previews, budgets, and authority boundaries are the foundation. The architecture distinguishes advice from execution and keeps explicit player orders ahead of delegated behavior.
LOCAL BY
DESIGN.
The default pipeline runs in the browser using CPU inference through WebAssembly. Once its models are downloaded and initialized, local commands do not require a remote model-serving API or a personal API key.
The model set is approximately 505 MB. Browser capability, startup time, memory, and cache behavior affect the experience. A microphone is optional; keyboard and mouse remain available. The current private package disables optional cloud-AI paths.
FROM THE OCTOBER ENGINEERING RECORD
Read the architecture story 160 / 160 held-out text cases · 98.8% speech-to-command in the documented noisy test · 0.36–0.60 s measured decision latency on the tested MacBook.
Recorded evaluation results, not universal accuracy or device-performance claims. The private build has not been certified across all browsers or microphones.