Three small models targeting different efficiency trade-offs: lowest-cost volume, reasoning quality at sub-4B scale, and multilingual breadth with a larger parameter budget. Llama 3.2 3B Instruct, released by Meta in September 2024, is designed primarily for edge and on-device deployment but widely available on hosted providers at sub-$0.10 per million tokens. At 3 billion parameters, complex reasoning and code generation are off the table, but classification, short-form summarization, and content moderation routing perform acceptably. The 131K context window is retained, useful for routing or classification over long documents. If your workload is volume-heavy and quality-tolerant, the 3B is worth benchmarking before committing to a larger tier. Llama 3 community license permits commercial use. Phi-3 Mini 128K is Microsoft's 3.8B parameter instruction model from April 2024, trained on curated synthetic textbook-quality data. The bet on data quality at small scale pays off: it outperforms several 7B-class models on reasoning and QA benchmarks. The 131K context window is unusually large for a sub-4B model, making it viable for extraction and classification tasks that would normally push you to a larger host. MIT license — no commercial restrictions. At this size, latency and hosting cost are the primary draws; complex multi-step reasoning and coding are still constrained by the parameter ceiling. Qwen 3 8B Instruct is Alibaba's general-purpose model at 8 billion parameters, with a 131K context window and notably strong multilingual performance on CJK benchmarks. It competes with Llama 3.1 8B on general evals while outperforming it on multilingual tasks — a real advantage for products with East Asian user traffic. Per-token pricing lands below $0.10 on most hosted providers. Released under the Qwen license with commercial terms. Pick Llama 3.2 3B for maximum throughput at minimum cost on classification and routing tasks. Pick Phi-3 Mini 128K when you need MIT-licensed on-device or constrained-budget reasoning with long-context support. Pick Qwen 3 8B when your application serves multilingual audiences or you need stronger instruction following at 8B scale.