Three small models, three different bets on how to maximize value per parameter. Llama 3.1 8B Instruct is Meta's mid-2024 refresh of the 8B tier, bringing a 131K context window to what was previously a short-context model class. It became the default general-purpose baseline for sub-10B comparisons, widely hosted across virtually every inference provider and carrying the Llama 3 community license with commercial terms. Mistral 7B Instruct v0.3 holds a persistent position in this comparison because of cost and compatibility: routinely under $0.10 per million tokens, with native function calling added in May 2024. Its 32K context lags behind both peers here, but the Apache 2.0 license and broad existing fine-tune ecosystem keep it deployed in production pipelines that predate Llama 3.1's arrival. General benchmark quality — MMLU, MT-Bench, instruction following — has been surpassed by both competitors in this group. Phi-3 Mini 128K represents a different tradeoff at 3.8 billion parameters. Microsoft's textbook-quality synthetic training data pushes its MMLU and GSM8K performance above several 7B peers, making it surprisingly competitive on reasoning benchmarks despite roughly half the parameter count. The 131K context window matches Llama 3.1 8B. The MIT license is the most permissive of the three. The limitation is open-ended generation: at 3.8B, creative and conversational quality gaps become apparent. Pick Llama 3.1 8B when you want the widest provider choice and solid all-around performance with 131K context. Pick Phi-3 Mini 128K when you need to minimize hosting cost and the workload skews toward structured reasoning or QA. Pick Mistral 7B v0.3 only if you already have fine-tuned adapters tied to its tokenizer or need Apache 2.0 specifically.