All three sit under four billion parameters, but they serve different masters. Granite 3.1 2B Instruct is IBM's smallest production model — two billion parameters with a 128K context window that puts it ahead of most sub-3B alternatives on long-document classification. IBM designed the Granite 3 series with enterprise compliance in mind: structured output, tool use, and data-lineage transparency under an Apache 2.0 license. At this scale it is not competing on raw generation quality; it is competing on throughput, hosting cost, and the ability to process long documents without a larger model footprint. Phi-3 Mini 128K from Microsoft brings 3.8 billion parameters trained on heavily filtered textbook-quality synthetic data. The training philosophy shows: it outperforms several 7B-class models on reasoning and QA benchmarks, which is the whole value proposition. The 131K context window is genuinely uncommon at sub-4B scale, and the MIT license removes commercial friction entirely. Expect weaker performance on complex multi-step reasoning and open-ended generation where scale matters more than data curation. Stable Code Instruct 3B from Stability AI was a reasonable lightweight code-completion choice at its January 2024 release but has been largely displaced. The 16K context for fill-in-middle tasks and the Non-Commercial Community License — which requires a Stability membership for commercial use — create friction most teams will not accept when Qwen 2.5 Coder and DeepSeek Coder offer stronger benchmarks at comparable cost under permissive terms. Pick Phi-3 Mini 128K for general reasoning and QA tasks where its benchmark-per-parameter ratio stands out. Pick Granite 3.1 2B for enterprise extraction pipelines where long context at minimal cost matters. Treat Stable Code 3B as a research artifact unless you specifically need a local fine-tuning base with permissively licensed Stack data.