FROM gemma3:270m

# Force raw completion mode with no chat wrapper.
TEMPLATE "{{ .Prompt }}"

# Keep decoding highly deterministic and cheap for inline suggestions.
PARAMETER temperature 0
PARAMETER top_p 0.8
PARAMETER top_k 20
PARAMETER repeat_penalty 1.05

# Bound context and output aggressively for low-latency autocomplete.
PARAMETER num_ctx 1024
PARAMETER num_predict 16

# Stop immediately at the first newline so completions stay single-line.
PARAMETER stop "\n"