This group shares a long-context threshold — all three expose 128K or more context — but land at different parameter counts and are trained for distinct use cases. Granite 3.1 8B Instruct is IBM's enterprise-oriented 8B model, expanded from 4K to 128K context in its December 2024 3.1 revision. The Granite 3 design prioritizes structured output and tool-use fidelity for enterprise extraction workflows. IBM's benchmarks show strong function-calling performance, which is the primary reason to prefer it over peers in this group for agentic RAG pipelines. Apache 2.0 license, primarily hosted on watsonx.ai with some Together AI availability. Phi-3 Mini 128K weighs in at 3.8 billion parameters — roughly half the count of its peers — trained on Microsoft's curated synthetic corpus to punch above weight on reasoning and QA tasks. It holds competitive MMLU and GSM8K scores against 7B models, which makes it the most cost-efficient option in this group when the workload is structured reasoning rather than open-ended generation. MIT license with no commercial restrictions. Qwen 3 8B Instruct has the broadest benchmark profile of the three. It covers multilingual CJK and Arabic evaluation sets where neither Granite nor Phi-3 compete meaningfully, and its 131K context window combined with below-$0.10-per-million pricing makes it the default recommendation for most general-purpose tasks in this tier. The Qwen commercial license is permissive enough for most deployments. Pick Qwen 3 8B for multilingual production workloads or general-purpose tasks where benchmark breadth matters. Pick Granite 3.1 8B when enterprise function-calling and IBM tooling integration are the priority. Pick Phi-3 Mini 128K when you need the lowest inference cost and the task is primarily structured reasoning or classification.