Gemma 2 2B IT, Llama 3.2 1B Instruct, and Llama 3.2 3B Instruct are all designed primarily for on-device and edge inference. All three were released in 2024, and all three are priced at the absolute floor of hosted inference — but quality and context trade-offs are worth understanding before choosing between them. Llama 3.2 1B Instruct is Meta's smallest model, released September 2024 under the Llama 3 community license. At 1 billion parameters, it is targeted squarely at phones and edge hardware where the 3B model exceeds memory budgets. The quality ceiling is real: on most instruction-following and summarization tasks it falls noticeably behind the 3B variant. Its primary use cases are latency testing at the smallest weight class, triage pipelines where a fast cheap initial filter reduces traffic to a larger model, or as an on-device inference baseline. Llama 3.2 3B Instruct, also from Meta and released September 2024, is a more practical choice for actual application tasks. It handles classification, short-form summarization, and content moderation routing acceptably. The 131K context window makes it useful for classifying over long inputs even though generation quality at 3B is modest. Sub-$0.10 per million tokens on several platforms as of early 2026. Gemma 2 2B IT from Google DeepMind is a 2-billion-parameter model released July 2024 under the Gemma license. On structured tasks like classification and named-entity extraction it runs comparably to Llama 3.2 3B, and the 8K context, while shorter, covers most edge inference scenarios. Self-hosting on cheap GPU or on-device deployment is where the economics favor it over hosted API calls. Pick Llama 3.2 1B for true memory-constrained edge hardware. Pick Llama 3.2 3B for the best quality-per-token ratio in this parameter class with a long context window. Pick Gemma 2 2B when on-device self-hosting in a Google ecosystem is preferred and the Gemma license terms are acceptable.