OLMo 2 13B and OLMo 2 7B are Allen AI's fully open Apache 2.0 models — no usage restrictions, full training data transparency. The size difference is the primary decision variable: 13B delivers noticeably higher quality on reasoning and knowledge benchmarks (MMLU ~63 vs ~58 for 7B) at roughly 1.6–2× the per-token cost. Typical hosted pricing runs $0.10–$0.18/M tokens for 7B and $0.18–$0.32/M tokens for 13B, making both among the cheapest options in their size class. Throughput scales inversely with size. OLMo 2 7B can sustain significantly higher tokens-per-second on a single A100 instance — useful when latency or concurrent request volume matters more than raw accuracy. Both models share the same tokenizer and training recipe, so swapping between them requires no prompt engineering changes. **Where OLMo 2 13B wins:** tasks that need more reliable multi-step reasoning, summarization of longer passages, or moderately complex instruction-following. The quality gap over the 7B is consistent on structured output tasks. **Where OLMo 2 7B wins:** embedding pipelines, rapid classification, or any high-QPS workload where cost and latency are the binding constraints. The Apache 2.0 license also makes it trivially deployable on-prem with no legal overhead. Pick [OLMo 2 13B Instruct](/models/allenai--olmo-2-13b-instruct) when benchmark quality is the tiebreaker and cost is secondary. Pick [OLMo 2 7B Instruct](/models/allenai--olmo-2-7b-instruct) for maximum throughput per dollar on simpler workloads — both give you full model weights with zero license friction.