Three sub-4B models that represent very different design philosophies: on-device inference, enterprise compliance extraction, and reasoning quality at small scale. Gemma 2 2B IT is Google DeepMind's 2-billion-parameter instruction-tuned model from July 2024, built primarily for on-device and edge inference. On structured classification and NER tasks it runs comparably to Llama 3.2 3B and IBM Granite 2B. The 8K context window is consistent across the Gemma 2 line and represents the primary constraint for document-length tasks. Economics favor self-hosting or on-device deployment at this parameter count; hosted API versions exist but are rarely cost-justified at scale. The Gemma license permits commercial use but is not OSI-approved — a distinction worth noting in enterprise legal reviews. Granite 3.1 2B Instruct is IBM's smallest production Granite model, a 2-billion-parameter model released as part of the Granite 3 series with an unusual specification: 128K context at 2B scale. Llama 3.2 3B and Gemma 2 2B both cap at lower context lengths, which gives Granite 3.1 2B a real edge for long-document classification pipelines where you would otherwise need a larger model. IBM designed the Granite 3 series for enterprise compliance scenarios and tool-use workflows rather than open-ended generation. Apache 2.0 license is straightforward for commercial shipping. Primary hosting through watsonx.ai; broader inference provider coverage is growing. Phi-3 Mini 128K is Microsoft's 3.8B parameter model from April 2024, trained on synthetic textbook-quality data to punch above its weight on reasoning benchmarks. The 131K context window at under 4B parameters is genuinely uncommon and makes it viable for extraction workloads that would otherwise require a 7B. MIT license with no commercial restrictions. Pick Gemma 2 2B for on-device inference and edge classification where 8K context is sufficient. Pick Granite 3.1 2B when you need Apache-licensed enterprise extraction with long-document context at the smallest compute footprint. Pick Phi-3 Mini 128K for MIT-licensed reasoning and long-context tasks where quality-per-parameter matters.