Granite 3.1 2B Instruct, Llama 3.2 3B Instruct, and Phi-3 Mini 128K are all sub-4B models designed for inference at minimal compute cost. They differ most sharply in their training emphasis: enterprise compliance and tool-use for Granite, broad edge deployment for Llama 3.2, and reasoning quality via data curation for Phi-3. IBM's Granite 3.1 2B Instruct, released as part of the Granite 3 series under Apache 2.0, has a 128K context window at 2 billion parameters — longer than Llama 3.2 3B's window by a meaningful margin for classification tasks. IBM designed the Granite 3 series for enterprise scenarios: structured output, tool-use, and extraction under compliance constraints. The Apache 2.0 license is the most permissive in this comparison, straightforward to deploy commercially without further legal review. Per-token rates are competitive on watsonx.ai. Meta's Llama 3.2 3B Instruct is the most accessible of the three, broadly available across every major inference provider at sub-$0.10 per million tokens on several platforms. The 131K context window matches Granite 3.1 2B in practical terms, making it viable for long-document classification even at 3B parameters. The Llama 3 community license applies; commercial deployment is permitted under Meta's terms. Microsoft's Phi-3 Mini 128K is a 3.8-billion-parameter model trained on curated, textbook-quality synthetic data that lets it outperform several 7B models on reasoning benchmarks at a fraction of the cost. The 131K context and MIT license are both clean. If the decision criteria is raw reasoning quality per parameter, Phi-3 Mini is the leader in this group. Pick Granite 3.1 2B for enterprise tool-use and extraction under Apache 2.0. Pick Llama 3.2 3B for maximum provider availability at the lowest price point. Pick Phi-3 Mini 128K when reasoning and QA quality matter most and MIT licensing is a bonus.