School-district public IT needs an on-budget classroom-help bot with tight latency targets. Which model class should they favor for that constraint?

School-district public IT with on-budget classroom-help bots and tight latency targets should favor a small language model for text help prompts. Small models are a primary Foundry planning lever when cost and interactive latency dominate and the task is text Q&A rather than vision. Matching model class to budget and latency constraints is as important as matching modality. A compact vision model selected for lower cost and lower latency on text help prompts fails by modality: vision models are built for image understanding, they do not generate interactive text help answers for classroom chat. A distilled speech model selected for lower cost and lower latency on text help prompts fails by modality: speech models are built for audio transcription and synthesis, they do not generate interactive text help answers for classroom chat. A text embedding model selected for lower cost and lower latency on text help prompts fails by task type: embedding models output vectors for retrieval indexing and ranking, they do not generate interactive text help answers. Exam caveat: 'small' must still meet quality bars for student-facing answers; evaluate with representative prompts before locking the SKU. Operational check: load-test p95 latency under expected classroom concurrency, compare token cost to budget, and reassess if future requirements add worksheet-image understanding that needs multimodality.