.env.example

# Local GGUF model cache.
LOCAL_MODELS_CACHE=~/models/synarmo

# Default number of suggestions.
SYNARMO_MAX_SUGGESTIONS=3

# Compose generation defaults.
SYNARMO_MAX_TOKENS=5
SYNARMO_MAX_SUGGESTION_WORDS=1
SYNARMO_TEMPERATURE=0.5
SYNARMO_TOP_P=0.95
SYNARMO_LOGPROB_POOL=24

# llama.cpp GPU layer offload.
# This example assumes an Apple Silicon Mac with one integrated Metal GPU.
# -1 offloads all possible model layers to that GPU.
# Use 0 for CPU-only systems or when debugging GPU issues.
# Positive values offload only that many model layers. This is not the number
# of GPUs; Apple M2 has one integrated Metal GPU.
SYNARMO_N_GPU_LAYERS=-1

# Native llama.cpp startup logs.
# Set to 1 temporarily to print model metadata, tensor types, KV cache,
# Metal/CUDA buffer sizes, and model size during model load.
SYNARMO_LLAMA_VERBOSE=0

# Model downloaded by the llama-cpp backend if it is missing from the cache.
SYNARMO_MODEL_REPO_ID=QuantFactory/Llama-3.2-1B-GGUF
SYNARMO_MODEL=Llama-3.2-1B.Q4_K_M.gguf

# Voice assistant. Browser is instant and uses the operating system voice.
SYNARMO_VOICE_BACKEND=browser

# OpenAI TTS is used only when selected in the Compose UI. Keep this key on
# the local server; it is never sent to the browser.
# OPENAI_API_KEY=
SYNARMO_OPENAI_TTS_MODEL=gpt-4o-mini-tts
SYNARMO_OPENAI_TTS_VOICE=marin
# SYNARMO_OPENAI_TTS_INSTRUCTIONS=Speak warmly and clearly.