Gemma 2 2B IT, Llama 3.2 3B Instruct, and Phi-3 Mini 128K are all sub-4B parameter models from major labs, each taking a different approach to squeezing useful capability into a small footprint. The meaningful differences are in how they achieve quality: architecture, data curation, or distillation. Gemma 2 2B IT is Google DeepMind's smallest Gemma 2 model, a 2-billion-parameter instruction-tuned transformer released July 2024. Google distilled it from Gemini Ultra training data, which helps it punch above its raw parameter weight on structured classification and extraction tasks. The 8K context window is consistent with the rest of the Gemma 2 family. The Gemma license is permissive for commercial use but not OSI-approved. Meta's Llama 3.2 3B Instruct, released September 2024, is the most broadly hosted of the three, appearing across virtually every major inference provider at sub-$0.10 per million tokens on several platforms. Its 131K context window is the headline spec advantage over the other two models here — useful for classification or lightweight summarization over long documents even at 3B parameters. The Llama 3 community license applies. Phi-3 Mini 128K from Microsoft is a 3.8-billion-parameter model released April 2024, trained on heavily filtered textbook-quality synthetic data. This data curation strategy lets it outperform several 7B-class models on reasoning and QA benchmarks despite its smaller footprint — a genuine quality advantage over both Gemma 2 2B and Llama 3.2 3B at the cost of a slightly larger parameter count. The 131K context window and MIT license round out a strong profile for teams running cost-sensitive reasoning or document extraction pipelines. Pick Gemma 2 2B for the smallest possible footprint on Google-ecosystem hardware. Pick Llama 3.2 3B for the broadest provider coverage at minimal cost. Pick Phi-3 Mini 128K when reasoning quality matters more than absolute size and MIT licensing is preferred.