An ultra-efficient 1.7 billion bit sovereign assistant, built on SmolLM2 and refined via SFT to bridge the gap between low-latency local inference and