Gemma 2 9B IT, Granite 3.1 8B Instruct, and Mistral 7B Instruct v0.3 are all sub-10B instruction-tuned models from mid-2024 that remain in active use. They are close enough in capability that the decision between them usually comes down to context window requirements, licensing constraints, and which provider ecosystem you are already in. Mistral 7B Instruct v0.3 is the Apache 2.0 baseline here. The 32K context window handles most summarization and classification workloads, and pricing typically falls below $0.10 per million tokens. The v0.3 update added native function-calling support, making it a complete enough foundation for lightweight agentic pipelines. For teams that need the cleanest possible open-source licensing for fine-tuning or redistribution, this is the most permissive and mature option in the group. Gemma 2 9B IT from Google DeepMind is a 9-billion-parameter model released July 2024 and distilled from Gemini Ultra training data. On standard benchmarks it runs comparably to Mistral 7B and Llama 3.1 8B, with Groq hosting offering some of the lowest latency numbers for this size class. The 8K context ceiling is a harder constraint than either of the other two models here. The Gemma license is permissive for commercial use but not OSI-approved. Granite 3.1 8B Instruct from IBM, released December 2024 as an Apache 2.0 model, upgraded context from 4K to 128K tokens in the 3.1 revision — the longest window in this comparison by a significant margin. IBM tuned it for structured-output and tool-use scenarios, and enterprise RAG pipelines that need to ingest long documents in a single 8B-class pass have a concrete reason to evaluate it. Primary hosting is IBM's watsonx.ai. Pick Mistral 7B v0.3 for the best Apache 2.0 general-purpose baseline. Pick Gemma 2 9B for low-latency short-context inference on Groq. Pick Granite 3.1 8B when 128K context at sub-10B parameter cost is the deciding requirement.