Three sub-10B models with sharply different value propositions. Gemma 2 9B IT from Google DeepMind is the production-quality option: competitive MT-Bench scores at this parameter scale, Groq-hosted with some of the lowest latency available for small models, and pricing typically below $0.20 per million tokens. The 8K context window is the same constraint as the rest of the Gemma 2 line — short-form classification, extraction, and generation tasks fit; document QA and multi-turn memory do not. The Gemma license is permissive for commercial use but not OSI-approved. OLMo 2 7B Instruct from Allen AI, released November 2024, is the research transparency choice. Weights, training data (Dolma corpus), and training code are all Apache 2.0. No other model in this size class matches its auditability. The tradeoff is practical: the 4K context ceiling is even tighter than Gemma's 8K, hosted provider coverage is thin, and on pure benchmark quality it runs slightly below Llama 3.1 8B equivalents. The audience is researchers who need a fully reproducible base for fine-tuning experiments, data lineage audits, or academic evaluation — not teams building production inference pipelines. Qwen 3 8B Instruct from Alibaba, released April 2025, is the strongest general-purpose production option here. It matches Llama 3.1 8B on English benchmarks and surpasses it on multilingual evaluations across CJK and Latin-script inputs. The 131K context window is the decisive technical advantage over both Gemma and OLMo. Pricing is competitive with Gemma 2 9B across major providers, and the Qwen commercial license covers production deployment. Pick Gemma 2 9B for latency-sensitive production tasks where the 8K context limit is acceptable and Groq throughput is a factor. Pick OLMo 2 7B exclusively for research workloads requiring full training provenance and Apache 2.0. Pick Qwen 3 8B for new production deployments where multilingual breadth and long-context support matter.