Sub-2B Models Find Real Jobs

The best argument for very small models is not raw capability, it's fit: they shine when latency, privacy, and on-device execution matter more than open-ended reasoning.

• –Edge automation is the clearest real-world fit; one commenter is already running multimodal Gemma-class models on Jetson hardware for home automation and function calling
• –Small models work well as routing layers, prefilters, and speculative decoding helpers, where they reduce cost without needing to solve the full task
• –They are useful for structured, narrow outputs like title generation, embeddings, zero-shot classification, and synthetic training data generation
• –In practice, teams should treat them as glue models in a cascade, not as replacements for frontier models on complex reasoning or long-context work